LIVE: How to best learn AI &ML in 6 months

hi friends a very good evening and thank you for joining in could you please confirm in the chat window if i am audible and visible clearly so that we can get started with the session in a few minutes again we'll start the session itself at 7 pm sharp so that everybody can join in the session because the schedule time is 7 p.m right again i have everything set up on my phone also just to look at all the comments um so that i can keep track of what's happening cool uh yeah so i'm visible and audible right anyway i'll start sharing my screen in a few minutes i have the chat window on my computer also so uh we'll we'll surely there is a little bit of notes that i made so that uh again i'll share the notes with you at the end of the session the notes itself will cover all the details on what you should focus on what are more important what are less important and things like that hey folks uh thank you very much for joining in today i understand this is year end for some of you and it's it's it's vacation time for others right we have a long weekend also i completely understand that and thank you once again for joining in the session cool uh we'll get started with the session itself by about 7 pm or so so i'll try to answer some general questions because the schedule time is 7 pm and we'll wait for everyone to join in cool yeah sure nabil is asking us me to share the notes i'll surely share the notes at the end of the session i'll share the detailed notes with all the links that i have in it and i'll post it in the description section as well as in the comment section of this video itself i'll do that hey happy new year to everyone i understand that 2020 has been very rocky for many of us and yeah so i i completely get that hope 21 will be better with the vaccine around the corner uh i mean i hope things will go back to normalcy so hemanth is asking a question about gans so he we covered basic foundational mathematics behind gans as well as a whole full-fledged code right so we implemented again from scratch using tensorflow gradients and so this is one of the sessions that we've done it's available in the course videos and in the live sessions in case you missed that please check it out uh again the most important thing about gan is the core concept itself as well as so we try to implement a cnn based gan basically an image generator right uh from scratch in tensorflow because that's the best way for you to learn as i will anyway tell you in this session so that's one thing of course there are more advanced forms of gans that we will surely cover in live sessions again the core purpose of the course that we have designed is to help students transition to careers in machine learning and ai i i completely get that most of you might already understand that so we are avoiding going too deep into one topic unless there is there is a there is a market need for that skills we cover gangs we even implement gans from scratch of course there are like hundred variations of gans and research papers we may not be able to cover all of it but in live sessions we are trying to cover some of these state-of-the-art techniques for example recently we did a live session on gpt3 and things like that so again folks please understand that we may not be able to again this is like one versus there are already 200 folks in the live session i may not be able to answer all the questions but i'll try my best to answer as many as possible and in case i miss any you can always email us at team at applied a course and we are there to answer your questions as best as we can cool so uh okay i'm just scrolling through the comments and i'm randomly trying to pick uh you want me to increase the microphone volume okay just give me a second uh okay so i think the microphone gain might be slower slightly okay i'll increase the microphone volume so those of you who whose volume was already high you might want to decrease it i just reduced the gain on my microphone so that it's easier for everyone to hear more clearly i hope this is better now [Music] okay so is the volume better now folks because some of you suggested me to increase the volume i increase the gain on my microphone is this better now so reva is asking can freshers like us get into earnest indian yeah we have students who are freshers even working professionals who joined instant young as data scientist roles i heard recently news that ernst young is trying to start uh hiring large number of earnest a large number of data science professionals in india and we already have a relationship with them because they've hired our students and those students have excelled so that's that's that i mean that's something that we are very sure we can do well on so cool cool so okay so ryan is asking an interesting question what do you say about fast ai course it's a good course no doubt about it i mean i think uh jeremy who is one of the founders of fast a course it's a it's a fairly well done course i mean i really liked it personally so i like a lot of courses again we learn a lot from other courses i learned a lot from how to teach better uh how to design the course better how to design assignments better how to do code dive deeps all of these we learnt from multiple courses and we are always looking for new ideas and fasta is a great source from which we learn how to do our courses better good okay good the volume is good so very good shravan asks a very good question in applied a course you focus a lot on python but some companies also ask for r that's i mean that's true shravan there are not every company would want you to use python there's no doubt about it but the fundamental limitation with r again that doesn't mean that the company says you have to only know r right so most companies if you know python they will still consider you because they understand that somebody who knows python can pick up r right it's again r is easier in some ways than python but the reason we focus so much on python instead of r is because python has more libraries and it's more widely used in production environments i'm not saying r is not used r and sas are also used but the proportion of companies is fewer okay that's that's the reason we focus more on python so arf asks the question arf khan which is best k n n or t sne for human action recognition neither of them i think i think you need to understand this human action recognition is a computer vision problem for which the state-of-the-art techniques are all based on uh convolutional neural networks right so that's that's one topic that we are going to do a live session on soon which is given i mean imagine i have a photo and i'm raising my hands like this can it recognize all the limbs all the all the joints in my body so that it can recognize my action so that's that mostly comes into computer vision of course uh it i mean state of the art in human action recognition is done using convolutional neural networks and not much of knn or disney like techniques um cool uh one second so somebody's asking i think i think somebody mailed us also about tiny ml it's a nice point i'll just cover that and we'll start the session right so uh so as far as anupam has a very good question so tiny ml right it's it's a very nice area again i like the naming because this name didn't exist just a few years ago tiny ml is all about taking machine learning and deep learning algorithms and placing them in low-end compute devices like for example imagine i mean nowadays smartphones have become terrific i mean they're almost as powerful as your laptops and desktops especially uh with apple a13a14 chips also called com865 i think 875 is also out but imagine you have an arduino board an iot chip right so there is a live session that we did i think it's called machine learning for iot it's it'll be there on our channel if i'm not wrong so you can check that where we talk about uh how you can take a machine learning or a deep learning algorithm and port it to a low end device like an arduino or a esp32 these are all microcontrollers which are are so these are all very simple cheap or raspberry pi right these are cheap microcontrollers which are used in iot devices uh so the best part of that is of course you can implement everything from scratch in c or something that's one way of doing it the other way is tensorflow itself you can take a tensorflow model and tensorflow is something called as tf light again we'll be doing a live session on it hopefully in jan or early february of next year or we're already in the end of the year so yeah of 2021 that's what i mean where we will take a tensorflow let's say convolutional neural network model tensorflow has this very nice method a very nice mechanism where you can take a tensorflow model convert into a tf lite model and deploy it on a mobile phone you can deploy it on arduino or any any iot device that's what tinyml is all about if you know core machine learning deep learning all of these concepts tidml is just a simple extension to it and thankfully of course understanding again if you if you're coming from electronics or electrical engineering background understanding microprocessor architecture microcontrollers computer organization certainly will help you design these things better right so but if you do not come from that background there are pretty good tools and pretty good libraries again you can write the code yourself in a low level programming or i can call it a low level programming language in a programming language like c right so that's possible cool so uh let's do one thing uh again the very interesting questions from people like b2 which is importance of mathematics so let me do one thing folks uh what we'll do is since it's 702 pm i'll i have some notes that i've written on what are the most important aspects so let me try to cover these probably i'll take about 45 to 50 minutes to cover these once i finish that the floor is open for a general discussion okay so is that i think that methodology is better because many questions that you are asking i would i would have covered them in in the notes that i have written okay then i'll surely come back we'll have a general q a okay as usual that that will be more beneficial because i can i can cover the broad topics first so let me start sharing the screen uh okay so i think everybody can see the screen now okay so let's go here so can you just confirm if you can see the screen just wanted to make sure that everybody can see the screen let me also see it on my phone just to make sure just to make sure everything is working i've just changed to my screen can everybody see my screen okay can everyone see my screen can you okay i think i can see my screen on my phone yes so you can see okay cool cool so let's get into the discussion and then uh we will finish the code the core discussion topic itself which is how to best learn ai machine learning in six months there is a reason why i put this number as six months because if you put consistent and persistent effort you can you can go a very long way in six months worst case nine months okay so we see a lot of our students who are putting in consistent effort from freshers to experienced professionals complete everything in six to seven months okay cool so let's let's go into it again this this whole lecture is not specific only to applied a course wherever you are learning a and machine learning from we wanted to give you a general guidance this is not just about applied a course or only our courses it's about general guidelines of course i will keep pointing you to some public live sessions that we have done pointers and things like that cool so first and foremost it's a very common question what are all the roles that are there and how are these roles different from one another okay it's a very good question first let's understand broadly there are four types of roles i'm not saying there are only four types of roles there are other types of roles also okay so broadly speaking there is this data analyst or a business analyst or a business analyst in data science type of role then there is a data scientist type of role then there is a machine learning engineer sometimes instead of a machine learning engineer you have a computer vision engineer or an nlp engineer or an ai engineer then there is an applied scientist sometimes these are also called as research scientists broadly these are the four categories of jobs that you have again some companies might have a slightly different nomenclature of naming things but broadly speaking these are the four categories that you will encounter now okay let me explain you how they are different so let's start with uh okay let's start with applied scientists again in terms of compensation applied scientists are some of the highest paid and for an applied scientist the expectation is again there are two broad angles on which any machine learning for any data science or machine learning role you will be tested on you will be tested on your mathematical rigor everything from foundational mathematics to advanced deep learning techniques and you will be tested on your coding skills okay so for applied scientist roles or for research scientist type of roles which are different from data scientist roles the expectation on your math front is high to very high again if you interview your top product based companies they would expect you to know the math inside out okay they would expect you to even modify the formulation of problems and work out when the problems are modified fundamentally at a mathematical level when i say math i mean the foundational math required for machine learning and data science roles as well as the math that is underlying most machine learning and deep learning techniques that's what i mean by this okay just let me check if there are any issues or complaints in the chat window okay okay so cool cool not no no complaints whatsoever so everything looks everything looks and seems clean okay cool so now coming back to this right so for an applied scientist you have to be very very strong on the mathematical foundations as well as the internal mathematics of every machine learning and deep learning technique and often times you are expected to be able to change the formulation and work out the details in an interview on the coding front the bar itself could be between low to medium so when i say high i mean it's really high high to very high sometimes so on coding front depending on your background imagine if you are coming from a non-cs background they would expect you to be able to write basic code right they would be they would i mean they may not expect you to know data structures algorithms very well but if you come from a computer science background or a software engineering background they would expect you to know the basic data structures and the hardness of questions themselves are low to medium okay so your interviews will focus for an applied scientist role more on math little significantly less on coding now the next type of role is called machine learning engineer this machine learning engineer is a role that is very very well suited for people who are already from computer science background or software engineers okay people who already have software engineering background or people who come from rit experience or people who come from computer science background for those people machine learning engineer is the most apt role because there is significantly more focus on coding okay so machine learning engineers are expected to know like multi-processing multi-threading core computer science concepts core software engineering concepts the math expectation for them compared to applied scientists is lower which means their expectation on the math is yeah you should know some of the stuff but we are not expecting you to know to the same mathematical rigor that an employee scientist is expected to know right so that that's so machine learning engineer role is mostly suited for people who are coming from computer science background or software engineering background and even in the interviews you will see much more focus on coding of course so in this probably you will have let's say four to five rounds of math one round of coding in this case you might have three rounds of coding and three rounds of math and machine learning when i say math i mean it includes machine learning okay then there is a data scientist role in the data scientist role the coding bar is low to medium again depending on your background if you come from cs background they'll ask you slightly high slightly harder questions if you come from non-cs background they'll ask you easier questions the math itself is slightly on the medium front so for a machine learning engineer and data scientist the math that is expected is roughly the same and it is certainly easier than the math expected from an applied scientist level right the coding expected for a data scientist is surely the expectation in terms of coding is less than a machine learning engineer and it's somewhat similar to what an applied scientist is now there is one more role which is data analyst for a data analyst the expectation in terms of code is low the expectation in math is between low and medium again even in terms of compensations at most companies your applied scientist typically gets the highest compensations so there are cases where machine learning engineers also get very similar compensations to applied scientists so one two three and four this is the typical order of compensations also right and again people can join i mean again you might ask okay for a non-cs person what are the best roles for a non-cs person the best roles would be data analyst and data scientist rooks for people who come from non-iit non-computer science background but if you come from non-csc backgrounds but if you are very strong in mathematics then applied scientist is also game right machine learning engineer role is something that is more well suited for computer science or software engineering professionals again these are general these are general hiring bars i'm not saying this is true for at every company for every role that's not true this is the general bar that we observe across multiple companies right but again we have seen interviews where data analysts are asked slightly harder questions than data scientists so again these are general guidelines that we are talking about in general the expectation from the interviewer or the company for applied scientist role in math is the highest encoding it is highest from machine learning engineers right so these are broadly the four categories that most jobs will fall into there are exceptions to this there is something called as a research scientist very few very few roles exist of that sort again sometimes the machine learning engineer can also be called as software development engineer machine learning okay so research scientists typically come with a phd in machine learning or in deep learning so that's very few roles also in industry of that sort right so broadly these are the roles and i'll try to explain the expectations for each of these roles in each of the topics okay i hope this gives you a good sense of what what is it that companies are expecting again this is this is a generalization this is like there could be companies which ask more code for data scientists than machine learning engineer i'm not so this is just a broad overview of roles that you will encounter cool so let's go topic by topic right so what is it that you should focus on in programming what what is it that you should learn on programming again very often programming is used as a way to measure your problem solving skills oftentimes most companies use programming again there are some companies which also have aptitude rounds right aptitude could be typically like your again the aptitude and reasoning questions could be anywhere from your 10th class mathematics level to cat level right again if you look at cat entrance examination right cat entrance examination has most problems which which require only 10th class math but more and more again the very few companies that have an aptitude around very very few literally a handful of them most companies test your problem solving and aptitude skills through programming because in a single so they might ask you a simple programming problem and if you solve it and write code they can test that you can solve problems and they can also parallelly test that you can write code again python is the most popular but there are companies which ask for r also but more and more over the years python has gained much more popularity because there are more libraries available so in terms of data structures you should know the inbuilt data structures very well no doubt about it and of course one more very important thing here is you should understand basic computational complexity right you should understand what is the time complexity what is the space complexity basic concepts we are not expecting you to be an expert at it but you should know the basic concepts of what is the time complexity what is the space complexity for the code that you have written and again if companies want to make the programming problems harder as you go more and more towards product based companies they would start asking questions which will involve recursion so within programming generally being able to write if else loops for loops right is is surely mandatory for almost every role but as the heart as as the level of the company increases as a company becomes more and more product based they will ask you more questions which will require you to solve it using recursion which is also called divide and conquer approaches in computer science right and again as you go more and more to product based companies they'll start focusing more and more on computational complexity one brilliant way to really become good at programming is implement inbuilt functions for example in python right there is this function called math.pow which computes x power y so you can give you can give two numbers x and y it will compute x power y now we all use it a lot in in python now my suggestion to you is if you really want to become a good problem solver or good if you want to become good at programming one of the simplest ways is look at all these inbuilt functions that you are using on regular basis and try to implement them from scratch how would you implement x power y it's a very good problem i used to ask this interview question at at one of the fan companies that i used to work at right i actually used to ask this like even even to computer science students right even to like really good cs folks from top universities of the world i would say in one of the programming questions i would say hey how would you implement x power y it's not trivial by the way right so similarly there is this function called like there is this very popular library called pandas which is used extensively in data science in pandas there is this function called merge which will merge two tables that you have now these are these are another very popular question that i used to ask at one of the fan companies which is uh how do you how do you merge these two tables right so a great way to improve your programming skills is to implement those functions again these are all our data science and mathematically related functions pandas dot merge again pandas is a brilliant library for data science applications now here your your your thinking on how to apply this merge function how to how to implement x power y so implementing these inbuilt functions will give you lot of exposure lot of practice also in addition to it you should know the basic libraries again there are thousands of libraries which are used in data science and machine learning and nobody can become expert at anything so nobody can become expert at everything right again even to become expert at one library is hard like i use pandas i use numpy i use matplotlib these are but i i can't say i'm an expert at pandas no if i get stuck i go right read the official reference manual of pandas and try to understand what's happening right so you should know at least the basic libraries numpy is for numerical computations matplotlib is for plotting pandas is for data operations so know the basic libraries you don't have to know everything but you should know at least some of the basic things that are most widely used again whenever you're writing any code many companies have the screening tests or coding tasks or coding tests one of the things that they're looking for is are you handling all the boundary cases right for example if you're implementing math.pow to compute x power y do you handle the case when y equals to zero are you handling this case or are you handling the case when x equals to zero and y equals to zero okay are you handling the case when y is less than zero are you handling the case when y is not an integer right so try to handle all these boundary cases in your code right your function may not do everything but you should handle these boundary cases gracefully okay so for programming the most important thing here is to write clean code write clean boundary cases the best way to practice i would recommend is to implement some of these functions as part of your practice okay again at applied a course we have some of these as assignments okay because that's one of the best ways to learn how to actually write good code right again we have a live session that i'm linking to here i'll share this whole document with you at the end of this session so this live session itself talks more about okay let me just see this live session let me click on this if i recall well uh okay so this live session here yeah so this live session is interactive interview session on python programming for ml and ai okay so we have done a full-fledged i think a two-hour live session on what sort of problems you can expect python related stuff in real interviews from services companies to top product based companies right again we try to cover various problems okay so again please understand that the best way to practice is to implement some of the inbuilt functions if time permits you can go to any of these popular platforms like lead code start with easy problems okay because remember except when you are a machine learning engineer the expectation in terms of coding is low to medium for all other roles okay in terms of coding so as long as you can solve simple easy problems and if if you can solve even a few medium level problems on lead code or geeks for geeks that's pretty good your your practice is fairly good unless you are looking for a machine learning engineer role where they might ask you more advice so for example imagine okay so i've put that case here also imagine you come from a computer science background or a software engineering background and if you're interviewing at a top product based company amazon google et cetera and you're interviewing for a machine learning engineer role which is more apt for computer science students or people with software engineering experience for those people or for even for experienced software engineers you are expected to know more than just a regular fresher or a non-cs student or a non-iit student okay there will be there could be one round which focuses more on data structures and algorithms sometimes the data structures they'll focus on will be simple stuff again even in machine learning we use data structures like trees and what is a decision tree addition tree is like sort of i mean one of the simplest variations of addition tree is a binary tree right similarly you have a hash table which is inbuilt into dictionary so there could be some questions on data structures and algorithms more so if you are applying for a machine learning engineer role at a top product based company and they would expect you to know concepts like multi-threading multi-processing etc if you if you come from a non-cs background they might excuse you for that but if you come from a cs background or if you're a software engineer they would expect you to know some of these basics especially at top product based companies cool so this is the expectation from programming now let's go to sql this is very very important it is critical for almost every role whether it's date again it's more critical for these three roles than for applied scientist roles for applied scientist roles i've been in interview loops where the candidate is a phd in mathematics we hire them even though he or she doesn't know sql that's okay but for data analyst data scientist and machine learning engineer roles this is critical you can expect a full-fledged dedicated round just for sql in reality i mean this happens a lot right so companies just have a dedicated round for this let me just check the chat window to see if there are any glitches oh okay so i'll come there i'll come there so so this is this so in sql itself what you should be able to write comfortably is nested queries okay if you can write nested queries slightly more contrived complex queries you are good if you really want to get a sense of what type of queries you should be able to answer we have this whole interview called sql for data science machine learning and sd and software development engineering interviews i'll provide the link here where we have covered multiple problems from easy to hard okay a simple rule of thumb is in sql can you write nested queries if you can you are you are all set for data science data data analyst data scientist and mle roles again remember especially for data analyst and data scientist roles sql is given much more prefer much more priority because you might say hey i am not as i'm not a computer science student but you are expected to know sql well right so sql is often used to test problem solving skills for data analyst and data scientist rules again very very importantly please don't think that if you join in a data analyst role you will stay in the data analyst role i've seen a lot of folks in my own professional experience who get promoted from data analyst to data scientist i've seen people who go from software development engineer to machine learning engineer roles i've seen people who go from software development engineer to applied scientist roles i've seen people who go from data scientists to applied scientist roles by building the gaps and skills that they may not have for example on on the math front if you're medium and you join a company as a data scientist you can improve your mathematical rigor and move into an applied scientist role through an internal review process or an internal interview process right so these are not fixed people move between these roles a lot in the real world cool so again sql is super important in sql the test is can you write slightly complex nested queries that's that's the key test that's that's a level you should you should be comfortable with next comes all of the data analysis and foundational mathematics a lot of people ignore this but this can be a break make or break thing in many interviews you will typically have one or two rounds of interviews on these topics oftentimes sometimes more especially for a data analyst and a data scientist role you will have more and more you might have just two or three rounds of just data analysis foundational mathematics etc in this plotting is super duper important i mean you have to know all the top types of plots how to clean the data how to pre-process your data if data has missing values how do you handle that so all of the data pre-processing cleaning plotting there are tons of types of plots of course you don't have to know everything but you should know at least the core ones what is a pdf what is a cdf what is a what is a qq plot used for some basic stuff you should know similarly probability and statistics i mean oftentimes i've seen a lot of companies which have a dedicated probability and statistics round or rounds right for example again we have a live session on this where we have solved a bunch of interactive problems in probability and statistics which is there on our youtube channel so for example somebody could give you a problem which requires you to apply bayes theorem can you solve it again i i worked in some teams where one of the lead interviewers would say hey if a candidate can't solve a base theorem problem is not ready for data science rules similarly there could be questions on hypothesis testing p-value right various hypothesis testing mechanisms non-parametric hypothesis testing and things like that so you have to be fairly comfortable with basics of data analysis most of the concepts in probability and statistics these are foundational you should know distributions you should know conditional i mean of course if you know base theorem you know conditional probability and things like that right so very very important you should know a little bit of non-parametric statistics also then comes data visualization again data visualization and plotting can be slightly interchanged but there are more advanced visualization techniques like psn etc that you should also know because they're more widely used in the industry similarly on the foundational math front you should know linear algebra again i've seen companies have almost dedicated round for foundational math where they focus purely on linear algebra right you should know equations of hyper planes you should know what is principal component analysis you should know what is singular value decomposition core mathematical concepts okay and the depth to which you have to know these concepts will increase if you are interviewing for an applied scientist role or for a data scientist row okay sometimes for some roles just an intuition of pca and intuition of svd might be enough but for other roles they will dig deep into the mathematics right you we co i mean this is super duper important i've been in interview loops where if a person performs well in these rounds we just hire them because they have the mathematical foundations to pick up machine learning and deep learning as and when they need it another very very important topic is numerical optimization okay suppose you have again this this boils down to basic minima maxima gradient descent stochastic gradient descent and all the sorts of algorithms right up to adam optimizer in your deep learning right so you have to get comfortable with all these optimizers which are most widely used in machine learning and deep learning of course non-parametric statistics is good to know if you know it it's great if you don't know it it's still okay for example non-parametric statistics the simplest stuff would be median percentiles quantiles iqr right simple stuff two more complex hypothesis testing using non-parametric statistics etcetera so again what i'm listing here is the bare minimum the more you know the better any day okay nobody will penalize you for knowing more but these are the bare minimum stuff that you should know now in these techniques or in all these concepts what are the stuff that you should focus on for example these are these are some of the most important things that you should focus on because we see some students who know these things but we can't answer these fundamental questions for example what technique to use where where do you apply this technique i ask these questions okay i often ask this so when you're learning about a technique or even a simple plotting thing like a probability density function or a or a cdf right these are very simple plot techniques i would ask hey where do you use a pdf and where do you use a cdf okay very simple questions and pros and cons of each of these methods from plotting techniques to statistical tests like some people say hey i'll use a t-test i'll say hey where does t-test fail that is nothing but the limitation or con of this test right some people say hey i'll use a chi-square test i'll say hey what are the advantages of chi-square test and where would it fail right those are very very important stuff that you have to know so you have to know where to use which technique what are the pros and cons of each technique what are the limitations of techniques you have to exp you have to be comfortable explaining the math for example if i ask you okay can you explain me about a metric like scale divergence in probability or i ask you can you explain me conglomerate spin off test okay or i can exp i can ask you can you explain me the underlying mathematics between t test or chi square test can you work out the math from first principles right because if you know the underlying math the chances that you will crack interviews is more this is one place where lot of lot of learners lack the if you don't know math you are not prepared for a data science career in the long run you might get in in the initial days but slowly you will not improve equally importantly my suggestion to you is because you have learned the basics of programming right when you learn a new concept try to implement it from scratch because that will that will improve both your programming ability and your conceptual ability for example let me give you a simple example this is a question that i used to ask at amazon myself okay so people say hey a very very simple question okay can you write a function to which i will give some data okay i'll just give a list of data can you write a function which will plot the probability density function i mean i mean this is non-trivial question right what we do a lot in practice is we'll just use matplotlib or c bond one of these libraries to plot the pdf but my question is different the reason interviewers ask these questions is because in one question i can understand whether you know the internals of pdf do you know the underlying math behind pdf and can you write code in one question i'm shooting two words i'm testing your programming skills i'm testing your underlying math knowledge also in a single question i'm doing both right so these are also very very popular questions and if you can implement a technique from scratch you will never forget it that's why it applied a course we in our assignments we make students implement some of these techniques from scratch we make them implement a logistic regression with regularization from scratch because that's the best way to learn it again very very importantly you should know where not to use a technique and this comes primarily from the limitations of each technique equally importantly you should know answers to why why is again you might say for example take anything right for example for for probability i mean take any concept in probability statistics right for example you have a t test okay i can say hey why can't t test be used in so and so situation or why can't chi square test to be used in so-and-so situation why is permutation testing required in this situation or why are you using a median and a percentile why are you not using a mean or a standard deviation or variance all this stuff so you will encounter a lot of questions on why you are using it why you are not using it why this is so these y questions can increase in complexity a lot right because these y questions can go to very very again you will be able to understand more and more of these wise if your underlying math is strong okay so whenever you are understanding the math ask yourself this basic question whenever you see any equation you should ask hey why is this eq why is this equation like this what happens if i change this equation those are very very very very important stuff that you should focus on especially in the foundational math and data analysis so now now comes classical so once your math once your foundational math is covered right once your foundational math and data analysis is covered then it all boils down to classical machine learning techniques again there are hundreds of techniques what i am listing here is the bare minimum that you should know okay these are bare minimum stuff that you should know i'll tell you what you need to know about each technique you should know of course k nearest neighbors logistic regression linear regression support vector machines both the linear version of it and the kernelized version of it basic decision trees random forest random forest is nothing but your bagged models then gradient boosted dish entries very important libraries like xg boost cat boost etc then how do you perform regression using decision trees how do you perform regression using svm how do you perform regression using knn regression using random forest all of that stuff basic clustering techniques k means hierarchical clustering the more you know the better always there is no limit to how much you can read okay i mean i believe that even after working in this field for close to close to what 12 13 years now what i know is just like less than one percent of techniques that exist in in research community right but these are the bare minimum you should know basics of matrix factorization you should know recommender systems these are all classical machine learning stuff that you should be very very comfortable with and in each again if you know more always better there is no doubt about it but these are basic stuff that you should certainly know again if an interviewer asks you a technique that you're not comfortable with you can say hey i don't know this but i know something related to this for example in all these techniques you should surely focus on the underlying math i cannot emphasize this enough if you know the math you can answer most of the y questions again when you are learning the math ask yourself why is this being done like this try to derive everything from first principles okay we tell this to all of our students at applied a course also for example let's assume you learnt logistic regression okay then you learned all the math okay very good take a piece of white paper try to derive everything from first principles which means don't make any assumptions try to derive whole logistic regression either using principles of probability or principles of linear algebra and optimization derive everything and at any point very likely you'll get stuck that's perfectly okay ask yourself why is this being done like this right for example why are we using l1 regularization why not l2 regularization okay again a very very popular question that i used to ask from top universe students from ph even phds from machine learning from top universities is hey we use l1 regularization and l2 regularization right why don't we use l4 regularization as much why don't we use l.5 regularization as much it's a very again this y questions you can make them very very very very complex as you want so always try to understand the math derive everything from first principles on a piece of paper because that will give you the mathematical rigor and depth okay if you have depth you can crack not just data scientist roles even applied scientist rules if time permits try to implement at least some of these techniques from scratch for example let me give an example right if time permits again k n is very simple to implement you should be able to implement it there are some specialized data structures related to k n okay especially if you're interviewing for a machine learning engineer role they might ask you about kd trees which is a very popular data structure used with respect to knns they might ask you about locality sensitive hashing so knowing machine learning specific data structures and algorithms is very useful again especially for machine learning engineer roles you'll get questions related to this try to implement at least one of these implement either decision tree random forest or gbt one of these okay again in one of the interviews when i interviewed at amazon in one of the interviews my interviewer said explain me how drive gibberity from scratch and he would stop me everywhere and say hey why are you using this why are you doing that why are you not why are you using sudo residual why are you doing that why are you not doing that all that stuff okay because he wanted to see because i used gbd in her other project that i did just before that he wanted to test if i really understand this inside out or not okay try to implement some of these things from scratch when i say from scratch don't use scikit-learn of course knowledge of these libraries is always helpful you should know these basic libraries whether it's xgboost scikit-learn more libraries the better always but try to implement it from scratch without using scikit-learn and other stuff because that will give you the confidence that you know the underlying math and you can change it right so that's very very very very important similarly you should know about bias variance tradeoff of each model what is the computational complexity of training what is the computational or what is the time complexity at runtime basic stuff you should understand about every technique similarly for each technique in the real world how would it behave imagine if the data has outliers how would this model behave or how would you make the model work if there are outliers in the data imagine if you have imbalanced data how does this model work what if the dimensionality of data is large would the model even work what are the best feature transforms you can do so that this model works well similarly what all metrics right again matrix is a very very important topic in classical machine learning what all metrics would this would this model try to minimize or maximize right very very important concepts these are again understanding the models inside out being able to implement them from scratch is very very important of course knowledge of libraries is important but so is answering okay what happens if your data is imbalanced and you're trying to train a logistic regression model or how do you fix it right so all these are fundamental questions that you should be able to answer again equally importantly for every machine learning technique you learn you should know what are the limitations what is the optimal scenario in which this will work very well what are the techniques you have to also know where not to apply this technique it's always good to know some slight variations of these techniques okay for example let me give an example you know logistic regression with l2 regularizer you should ideally know logistic regression with l1 regularizer also okay knowing at least one or two variations of these techniques is very very helpful suppose you know k means do you know k meteorites simple variations are good enough equally importantly try to apply them on real-world data there is so much of real-world data available on platforms like kaggle just pick up a problem from kaggle and just try to apply all the techniques you are learning using libraries or your own implementations on this data that will give you a lot of clarity on how to apply them so what you need broadly is you need foundational understanding of the mathematics you need to be able to implement it you need to be able to understand things like computational complexity bias variance trade-off or overfitting under fitting what happens in various real-world situations and the limitations and variations of techniques and where to apply them and how to apply them in the real world these are very very important for most again for data science machine learning engineer and applied scientist these are mandatory okay similarly deep learning is a vast field right there is there are literally thousands of architectures but there are basic stuff that you should know in late 2020 early 2021. again this changes from year to year as more techniques become popular for example now you should certainly know basic multi-layered perceptrons back propagation algorithm like i've been in so many interview loops where the interviewer just asks about back propagation right just basics he'll take a function and say hey how does back propagation work with this okay similarly you should know convolution neural networks everything from basic alex net at least up to rest nets if you know more models that's better right this is the bare minimum that you should know when it comes to recurrent neural networks you should know lstms gru's bi-directional lstms of course you should know auto encoders right all this how word to vect works internally similarly attention mechanism is becoming more and more popular both in nlp slowly in computer vision also you should know transformers button gpt models and gans right at least this is the bare minimum you should know the more you know the better it is no doubt about it there are some specialized architectures also like for one for one shot learning and all that if you know them it's great but at least this is the bare minimum that you should know okay again for each of these techniques you should know all this the underlying math try to derive it from first principles implement from scratch learn libraries again there are broadly two three libraries that are very popular in deep learning right so there you have so for example you have tensorflow right you have tensorflow again keras is more or less uh incorporated into tensorflow then you have pi torch right these are two broad libraries at least no one if you know one picking up the other is fairly straightforward suppose if you know pytorch learning tensorflow is very easy because you have the conceptual understanding of things if you know tensorflow pytorch is again very easy to build at least no one knowing more always helps but remember becoming an expert in all of them is also very very hard takes time in six months we can't pull it off again some of you might be thinking hey how do i learn all this in six months i'll come up with a timetable for you a crisp clean time table just give me a few minutes now again you might say hey i want to pick up specialized topics like computer vision right so if you want to if you want to excel in computer vision again most modern computer vision is deep learning focused there is non-deep learning focused computer vision also i'm not denying that but state-of-the-art computer vision is all deep learning focused using some variations of convolutional neural networks for example take image segmentation or take object detection or take image generation using using computer vision based gans right so if you are looking for a career in computer vision you should know at least the basic types of problems in computer vision and various techniques for this for example for object detection you should certainly know yellow v1 v2 v3s right because you want to focus on computer vision you should certainly know these basics if not everything else right similarly if you're looking at segmentation algorithms you should know things at least the basics like units the more you know the better it is at least the basics you should know right similarly if you're looking for nlp again there are if i have to pick up pick one sub area of machine learning and deep learning where there are most roles it is nlp there are fewer roles especially in india with computer vision but there are more roles with nlp you should certainly know basic vectorizations you should know basic pre-processing clearing of text data you should know what to back how it works internally right you should certainly know more advanced uh featurizations of text like birth and of course the internal architecture of word because that's what is being used more and more and more again word to vec used to be the used to be the sweetheart in nlp now slowly it's all moving to transformer based featurizations okay of course you should know basic stuff like tfidf all that you should know i'm covering all of that in pre-processing these are the basics that you should at least know for a career in nlp now if it comes to productionization productionization there is no end to how much you can learn but the bare minimum you should know is how do i productionize my model using a simple api how do i deploy an api at least in one cloud-based system you pick any one of them you pick aws azure pick any one of them again it's impossible to be an expert in everything for example i have worked on aws extensively because i worked at amazon i did not work in azure and gcp if i go to an interview and somebody says hey i worked on them briefly but i'm not an expert at them in any way so if somebody says hey how do you deploy your api in let's say google compute platform which i've used the least i've used even azure to some extent i would say hey i know how it is done in aws i'm sure there will be some equivalent in google compute platform right so knowing at least one of them or even some platform as a service like herroku right so at least knowing one of them is good enough right at least knowing one of them you don't have to be expert in all of them because it's impossible for anyone to be expert in all of them i've seen engineers with like 10 20 years of experience who can't be experts in all of them they can only be experts in some of them right so these are the bare minimum you know of course if you want to learn more okay there is all this uh you there is a docker kubernetes right and there is whole of this uh there is all of this ml ops right which is basically devops plus there is a new role that is coming up which is devops plus or devops for machine learning okay where again this role this sort of role is most appropriate for people who are already currently devops who learn machine learning who understand the machine learning life cycle who understand about machine learning training right so for people who are already devops engineers if they can pick up machine learning there is a new role that is coming up becoming more and more popular called mlaps but there is no end to it trust me there is no end to productionization because this is constantly evolving field because i can talk about productionization on let's say an android device there is no end to it but this is the bare minimum that you should know bare minimum stuff you should know basically how do you translate your code that you have in ipython notebooks into apis how do you deploy them basic stuff you should know most importantly try to build a portfolio of projects my suggestion is at least have two or more projects try and solve two or more problems you can pick data sets from kaggle you can collect your own data sets but try to do them end to end when i say end to end everything from data acquisition to productionization of the model on a cloud-based platform this will help you enormously and amongst the two or more problems that you want to solve you can get tons of problems from kaggle don't worry about it from every field there are data sets that you can publicly access and build something on it my recommendation in 2020 2021 as we are nearing 2021 is at least pick one problem with nlp that's what we strongly recommend because the need for data scientists and machine learning engineers who know nlp is only increasing because this is the most widely available data that we have okay at least one problem with nlp rest of them you can choose some problems from your domain of expertise etc very importantly please write a detailed blog this again i i'll show you i'll show you some again we have done a live session which i'm providing a link here on a bunch of portfolio projects that you can do based on your experience we discuss about what types of problems you can solve as a college student what types of problems you can solve as a tenure experienced professional right and in these problems suppose imagine you are using a deep learning model okay let's assume you're using yolo v3 i'm just i'm just making this up okay if possible at least implement one such algorithm from scratch instead of using a pre-built code the reason i am recommending this is because through your blog and through your portfolio you can showcase again what is the purpose of this portfolio to showcase your skills to recruiters try and implement at least one but as much as possible from scratch instead of using yolo v3's existing code that the others of yellow ev3 gave you it's a very very nice model try to implement it using tensorflow or pytorch whatever library you want try to implement it from scratch because then you understand how this algorithm is working internally and we have seen lot of students who have done this who have cracked interviews left right and center because this pro this portfolio project will be there at the top of the resume and the interviewer says hey interview will say something about yellow v3 then the candidate will immediately say hey i implemented yellow v3 from scratch here is my blog here is my code here is my github profile okay so i talk about it extensively in this live session which i've already done in the past i'll link to i'll share this document where you'll get all these details now you might wonder hey how do i do all of this in six months okay again i'm assuming again this is from our own data we see a lot of students completing in about six to seven months of course the most important thing here is the effort so we see students who are average not brilliant not below average the average student who puts in anywhere from 12 to 15 hours a week roughly two hours a day give or take those students this is a rough timeline that we see right python and sql maximum one month again don't try to become don't get stuck in any one concept keep moving because while you learn the basics of python and and learn some of these you'll get lot of python python practice when you're implementing the classical machine learning algorithms from scratch when you're implementing some of these math techniques from scratch don't get stuck to one thing keep moving that's very very important we see some students who are like hey i finished python but i'm not an expert you will not become an expert in one month once you learn python you will become you'll become better and better at it as you're implementing lot of techniques that you learn in data analysis machine learning deep learning and while you're building your own portfolio you're going to use python everything so everything that you're going to learn in python and sql you're going to use it throughout the rest of the five months right so effectively you're not learning python for one month you're learning it for all the six months similarly all the data analysis and math that you learn you can finish this comfortably in about a month but again these two are very very important i'm giving more time to these two topics because in the initial days this can feel slightly overwhelming happens to any one of us imagine if i were to learn let's say some complete if i were to learn let's say some pick a topic right suppose if i have to learn economics where i have no i know nothing i didn't study economics beyond my 9th 10th class suppose if somebody tells me hey let's start let's study economics or macro economics i don't know anything in it when i start learning in the initial topics i will be slow obviously because the initial topics are fundamental they're completely sometimes new to me so that's why i'm allocating more time for these initial topics because you might start slow but it is a persistent effort of 12 to 15 hours that is important if you if you are comfortable with python if you are comfortable with the foundational mathematics and data analysis rest of them will become much much faster and easier right so once once you get comfortable with again we see students who are finishing this in 20 days even 15 days but give yourself a month because these are initial topics they might take more time for you to learn from then classical machine learning algorithms once you get a hang of it right it becomes easier for example probably learning logistic regression is harder but once you learn logistic regression learning linear regression will become easier okay learning decision trees might take little time because it's slightly new but once you learn nation tree is learning random forest you can do it in a day or two all right so don't get bogged down by the number of techniques because once you get used to that thought process again these first two months or less less than equal to two months will give you the mindset and the thought process for you to progress faster in deep learning right because you learned all the classical machine learning techniques deep learning is all about using slightly advanced architectures and advanced optimization techniques that's all so you can actually once you get a hang of how to understand an architecture picking it up is very very quick right you can you can learn literally one architecture a day if the moment you get comfortable with understanding deep learning architectures okay again i'm not saying this will always happen even if you're learning one architecture in three days that's okay learn 1015 architecture you're good right then again very very importantly don't wait for you for yourself to learn all these to start your portfolio once you've covered a bunch of classical machine learning algorithms maybe a couple of deep learning algorithms start thinking of building the portfolio parallelly but maximum give yourself about six to seven weeks to build your portfolio again this portfolio building is not to be done at the end you can start it as soon as you learn a few classical machine learning techniques and work on the side okay again take my word for it the moment you become comfortable with programming and the moment you become comfortable with the math rest everything will fall in place your mind your thought process improves the first day you learn a new programming language or a new task it's always hard please don't give up on your journey be persistent promise yourself in the new year i mean i know that most new year resolutions never work but at least be truthful to yourself and say hey i will at least spend 15 hours a week again if you think you're an average student 15 hours 12 to 15 is good if you're an above average student you can do it even with eight to ten if you're below average of course you'll have to spend more time that's that's that's more a realistic estimate again this is doable we have seen students we have done faster than this also but this is a very very doable thing cool so in a nutshell what i wanted to tell you is this when you are learning machine learning data science everything understand the underlying math and know why you are doing something that why is very important behind every equation behind every plot understand why you are doing it try to implement everything from scratch because it will improve your programming but it will improve your programming insights it will also give you deeper understanding of the concepts like you can go to some of the top universities go to stanford for example at berkeley look at their machine learning and deep learning courses look at their exams they'll ask you to implement something from scratch because these professors who have been teaching for decades they also understand that once a student implements from scratch they understand the concept internally this is not something that i invented i've learnt it from other top universities of the world similarly understand all the cases pros cons where does it work what happens if there are outliers in the data what happens if i have imbalanced data all these cases for everything try to understand the cases most importantly apply the techniques that you're learning and solve real-world problems by picking by picking problems from kaggle and build a portfolio to showcase your work because if you don't showcase your work recruiters and others don't know what you have done in a nutshell these are the five things that i will ask you understand the math know why you are doing it implement things from scratch if you can as many as possible understand pros cons limitations various cases of everything apply them to real world problems and showcase your work you are ready it's as simple as that okay so again we have done earlier live sessions you can just google search for interactive interview questions there are tons of interview questions that we have discussed to for you to get a sense of it right again i'll link to this but let's open up the these are some of the concepts that i wanted to cover i'm sure i could not have covered everything so let's open up the discussion general discussion i'll try and spend as much time as possible uh okay so cool so now i'm back cool uh okay sounds good so let's tackle some of the questions just give me a minute i'll just take a water break and we'll get started cool good so bitu asks a very good question best books for machine learning here are some of my favorite books okay so for uh for python programming uh i like the python official documentation it's very good just go to python.org if i'm not wrong their tutorials are terrific that's one place to learn python sql w3schools go to any online resource there are tons of resources or you can because book any standard orale book about sql that's also will do again there's an orally book on python i i forgot the name of the book but that's also a good book or there's a book called automating boring stuff in python that's also a very interesting book in python similarly for data analysis and probability statistics there is this book called think stats it's a very simple very applied book from orelli if i'm not wrong that book is freely available online so pick up that book think stats very nice book machine learning there are two books that i really like one is called pattern recognition by christopher bishop and uh elements of statistical learning these are these are like two of the best books that i know of these are used in almost every university as standard textbooks for deep learning there is yan goodfellows very nice textbook but again deep learning is so fast moving that i tend to start reading technical blogs or i tend to read research papers but that book is a very good foundation book these are some of the books that i would recommend in general okay cool uh how much math should you know you should be able to so that's a good question gunja om prakash so you should be able to derive any technique from first principles on a whiteboard or on a white piece of paper that's the level of math that you should know ideally i mean if you know that you're safe you can even crack top product based companies if you know that level of math like for example i could ask you okay can you derive linear regression from first principles can you derive everything tell me all the assumptions you are making why you are choosing this model why this why that let's derive the whole damn thing if you can do it that's the level of math that you should ideally know okay cool uh so uh suffer says tell me about how to build a profile again on our youtube channel there is a live session called so let me look it up again i'll provide a link in this doc so this doc has a link on how to build a portfolio i'll share this doc with you right after the session and you you will be able to see that so there is a dedicated to our session on what sort of problems to solve based on your experience based on your background it's a very detailed two-hour session on how to build a machine learning portfolio of projects for career transition that's that's the title of the live session you can find it on our youtube channel it's publicly available so cool uh hanuman says how to answer if an interviewer asks explain cnn it's a very nice very very simple question it looks very simple you can talk about why we need cnns why can't we just use ml piece first very important what is a convolution operator why does it why are we using a convolution operator and why does it improve the results as compared to a standard multi-layered perceptron you can leave there and you can say hey this is the basic intuition or basic reasoning on why we have a cnn if you want me to go into more details i can tell you about various operations we have in cnn like max pooling like so you can you can keep going into the depths of it but give a give a quick two three minute overview on why we need it what are the key components of a standard cnn and what operations do they do like the convolution operation etc and ask them do you want me to go into more details that's good enough so uh okay so there is a question about uh the the the online one-year diploma with uh right so it's a very good question arvin that you asked so uh this program is starting in january um um so this is a program we are very very fortunate and we're very excited to do this one year diploma in a and machine learning with university of hyderabad so the fundamental difference between our applied ai course and the one year diploma in a machine learning would be the diploma machine learning ai is co is co-conducted by university of hyderabad and us so whatever i talked today will be there in that course without doubt but there will be slightly more academic rigor so we'll go into more mathematics right then i mean in some concepts if not all the concepts because there are some very senior university professors who were like dean of mathematics department etc who might be taking some classes there are also industry professionals who will be co-instructors in that so as an academic course it tends to be slightly more mathematically and academically rigorous so you'll have standard exams all of that stuff but in terms of coverage i think what we'll have is very similar to what we have in the applied air course will again the the the reason uh is conducting this diploma in collaboration with us is because what we bring to the table is the real world applied aspects so i am the i'm the primary instructor for all the courses but we have co-instructors who will be taking some sessions so we will continue to maintain the academic rigor of any university program along with all the applied aspects that we cover at applied a course but all the grading everything will be done by university of hyderabad through midterm semester and type of exams so that's the fundamental thing you will gain you'll you'll gain the knowledge in both places whether it's applied a course or pc or the or the online diploma course again remember the online diploma course is a purely academic course with one year so the applied ai course you can finish it in six months nobody is stopping you so one student asked why are you why is the diploma program one year why not six months because university grants commission which is part of which is called ugc which is under government of india has stipulated that all diploma programs should be at least 12 months anything less than that it's not recognized by ugc or government of india so we have to stick to their norms and processes because we want this to be both academically rigorous and recognized by government of india right so that's the reason it is so applied a course is more self-paced you can finish it in four months six months eight months whatever time you want the diploma program is more academically rigorous more academically focused i should say in terms of mathematical rigor right even in applied a course we cover most of the mathematical rigor that is there so it will be more academically focused cool uh uh one second okay once didn't understand uh uh so naman says i want to do a master's in ai and machine learning this is beneficial to do the course yeah surely i mean we have a lot of people who who pursue ms in us or mtech at iit's iac we have a lot of students who are masters degree students at iit's iasc even undergraduate student at some of these places where course participants and the reason they take the course some students even take the course before they start their ms program because this will give them the foundations to perform very well in their coursework and get into internships quickly right so a lot of students at indian institute of science iits etc take our course because it helps them because most university courses tend to be academically focused more theoretical focus what they learn from our course is in addition to the rigorous theory they also learn the applied aspects which is what benefits them when they go for internships or full-time roles cool uh so trojan has a question which is how to pick a problem statement and perform end-to-end uh operations on it so that's fairly simple right so for example again just watch that live session that i'll that i'll point to i think let me just look it up it's called okay just give me a second i forgot what it's called it's called uh applied a course um portfolio just a second uh projects i think it's so okay i just have to search huh so it's called machine learning projects for your career transition based on your current role okay so that's okay so let me just mute this yeah that's that's the title so let me just share it with all of you just a second folks um yeah so let me just put it here yeah i'm just pasting the youtube link of the two hour session that we have done about how to pick a problem what's again based on your expertise based on your experience based on your educational background and how to solve it end-to-end end-to-end means everything from data acquisition doing data analysis doing statistical tests doing feature engineering building machine learning models coming up with business problem the metrics that you care about doing that whole thing productionizing the model deploying the model maintaining the model retraining the model that whole thing that whole data science life cycle we have another video on our youtube channel it's called uh data science life cycle just check that that will tell you what do i mean by end to end that's very very important to think through right cool okay so let me just scroll down okay so sriram has a good question so sriram has i like the question so he says there are roles like full stack data analyst how good are these roles is it good to join some startup companies as full stack analyst again this title is borrowed from full stack development right in software engineering so where people develop both front-end javascript back-end servers so when somebody says a full-stack data scientist or a full-stack machine learning engineer they're saying hey we don't we want somebody who can do everything from data cleaning to model productionization that's what they mean in a nutshell that's what again it's just a fancy title what it means is sort of like a machine learning engineer if i have to put that into one of the roles it's closer to a machine learning engineer or sometimes even a data scientist because data scientists do productionize models right so a full stack data scientist or a full stack machine learning engineer basically means somebody who can do everything from data analysis data cleaning data acquisition obtaining the data scraping the data again we have another live session it's called how to scrape data from websites using using beautiful soup and other libraries right so how do you do that up to productionization of models using apis you're basically microsoft business architectures etc that's what they mean uh so rohit says okay so let me answer this question also rohit says i'm deciding to again folks i'm very sorry i'm not able to answer every question i'm just scrolling as fast as i can pick one question and try to answer it i'm just doing it more randomly here okay so cool so uh rohit had a question which was uh i'm planning to do the theory first and then the assignments in applied a course we have students who have we have done that successfully there's nothing wrong with it but we would say do them do them like in the order that are there that's sometimes beneficial right i'm not saying that's the best way that way you'll get a hang of things instead of just binge watching them but we have had students as you might have seen in some of our success story videos right some students just binge watch like netflix series and then finally they sit down and write code there's nothing wrong with it that's your call okay we will leave the addition to you so abdullah says what algorithm are you applying to pick a question random okay so uh there is there is one very nice question that just caught my eye and i missed it okay so this is shreyansh he's saying his knowledge of framework like django flask also required in the field you don't require to know django jango is like a big full-fledged back-end platform right you don't have to know django even if you know how to build a simple api simple api using flask or using a fast api even using django no problem with it if you know how to build an api how do you if you if you can expose your model through an api that's good enough you don't have again django is a massive platform i can't claim i know django i know some parts of django because we've used it for some components of our course but i don't know django fully it's only foolish to say if i know django right so i know again it's very important as engineers right for us to pick up whatever is needed when it is needed for example let me give an example right so let's take django so i know some parts of django i know how to build an api in django because we've used it for some components of our website there are some components that i don't know if there is a need for me to learn it i will just read the documentation because i know python programming i know the basics of how apis work how how basic internet works so i will just read the documentation pick it up and that's the skill that we all need to build right so uh so heyman says how to keep practice on deep learning frameworks uh i think if you mean how to become good at tensorflow or pytorch take a model implement it from scratch take a take a gan simple gun or take a transformer model again we did this in one of our live sessions where we took a transformer implemented a transformer it's fairly it's fairly like cumbersome takes a lot of time in terms of code also i think we did we did a live session for about four hours two two hour sessions where we implemented a transformer in tensorflow from scratch and it was a terrific experience because first you understand how a transformer works internally it's also a great practice for uh for for tensorflow code like it was a great great practice for me also when i was preparing for these live sessions i said okay let's let's do the whole thing in tensorflow i mean it was a good again the best way to practice any deep learning framework is very simply pick a model implement it from scratch nothing beats that cool uh okay so arun has a question he says i have 10-year experience in mainframe and java support can i transition to the ml rule yes so what you should certainly look for is again i was just i mean i just let me give you a context i was interviewing one of our students just now i mean just i think one hour before this session i mean i was just talking to him and he said he has seven years experience in technical support okay he's a senior technical support engineer so technical support not even coding okay he has not coded extensively he finished the course successfully in one year now he's joining here as a in a as an assistant manager for a data science team so he has successfully transitioned so we have and he's he's a bsc math with mca if i'm not wrong with seven years experience so there are people like that with seven even more years of experience sometimes even from non-technical fields or somewhere near technical but not exactly coding and all that stuff so it's very possible it all requires you to put more effort than somebody who already knows coding that's that's the key here you can learn you can certainly learn that so ashok if you have 10 years experience in mainframe or java support you have to brush up you have to become again i'm sure that as as a mainframe engineer you have written cobalt code you've written i'm sure general code so if you have the coding mindset and if you are willing to put in the effort i think you can successfully transition no doubts about it uh one second so uh one second what is this question i i just moved too quickly one second so i am a second year student in b tech should i start from applied ai shiva my suggestion to you is uh you are currently in second year right my suggestion to anybody who is doing your b tech right b tech first year second year try to become good in programming become good at data structures algorithms because the moment and of course become good at the math that you learn in engineering mathematics you can start it in second year but third year is also good time for you to start again if you start in second year or probably in second year second semester as a cac student i am assuming you know basics of programming try to become good at programming try to become good at basic data structures algorithms then taking the course probably in a second or second semester or third year first semester is a very ideal situation because then you have the basics you can move very very fast in the course and strive for internships right after your third year that's that's a nice that's a nice path that you can follow uh so ryan says has pytorch taken over tensorflow and keras i mean what do you mean by taken over that's that's uh again pytorch is very very popular with researchers so if you look at lot of research publications at top conferences journals more people use pytarch but if you look at production environments i think they're they're they're neck to neck right there is no again let's not get too caught up on a platform or a framework or a library because they come and go trust me okay we have tensorflow pytorch today tomorrow there will be new thing what is important is pick any one of them all of them are fairly good fairly mature right now for example in pytorch can you train a model and deploy it on an android phone it's i'm not saying it's impossible it's much harder than do get in tensorflow okay again remember tensorflow is being used by companies like google and maintained by thousands of open source contributors i'm not saying pytorch is bad or pie torch is good pick any one of them that's perfectly all right pick any one of them right now in the research community pie torch is more popular but in production they're neck to neck and that's perfectly okay nobody can be perfect at every library trust me like it's only foolish to think if you know numpy everything inside out no i've been using numpy for almost 15 years i can't claim that i know everything in numpy i only know the functions that i need that i use if i forget something i go read the reference manual right same with this if you know the concepts picking up a new library is is just a weekend's work or suppose tomorrow there is a new library because i know about the computational graph i know about the concepts in tensorflow for me picking up a new library is just a weekend's work okay cool uh is is is the placement scenario different for diploma and applied a course yes so uh the the at applied a course we have the job guarantee or money back program because we want to be completely aligned with student success in the case of the diploma with university of hyderabad because it is under the ugc under government affiliates a government research university by the way so they will not have a job guarantee or money back program as far as the placement assistance is concerned it will be exactly same the same placements team that we have we are just we are just increasing the manpower hiring more experienced folks hiring new folks hiring more hr folks to help us with with with our with our placements team so it's only going to get so everything that we have in the applied a course in terms of placement assistance will be there for students who graduate at the end of their uh of course i mean the most important thing here is we expect you to have good grades good marks in the uh diploma right please try and if you're joining the if you've got an admission letter again i think the university started sending admission letters for the last couple of days or so as in when we get an admission letter we are forwarding it to all of the students please commit yourself to putting in those 12 hours of effort every week right because that's very very important again remember university of hyderabad has a very rigorous academic credentials so the exams will be conducted by them we will give you all the mentorship all the education you need but you have to perform well in their midterm and semester and exams right so as long as you perform well there is nothing wrong with it it's great for you okay very interesting so sagar has a question it's a nice question is it good to do an ms in data science and ai from foreign countries to get a high-end good package again one of the biggest reasons why indians go and pursue ms in the us is not because they're passionate computer scientists are passionate about computer science because an ms in us in any field gives you the the the level what you call it the permission from the government to work in the us for the first three years or two years whatever that is called right so surely if you're pursuing if you're going to going to us because there are more job opportunities and the compensations is high obviously you'll get more comp no doubt about it right i mean just in pure dollar amount term cool uh so niraj has a question can we do a live session on self-supervised learning sure that's a topic that i'm very much looking forward for very interesting topic we'll surely do that we'll sure again parts of it have been covered in other models that we did but we'll surely cover that okay uh ankita has a very nice question i really like her question so she has about five years work experience in etl right like standard database pipelines etc what is better for me a data architecture role or a machine learning engineer both roles are good for you actually to be honest with you again a data architect role is a easier transition for you to be honest with you because it's what it's it's very similar to what you've been doing from an etl engineer with five years experience data architecting is like a it's like c and c plus plus right a machine learning engineer role is is is as good as a data architecture role but with more learning that you have to do so it's a career choice that you have to make do you want to be in data architecting role in in for the next five years or so or do you want to move to machine learning again if you are good at it in data architecting you can get the same compensation as machine learning engineer again at the end of the day in any career path you take it is your excellence that matters i have seen software engineers who are brilliant like literally brilliant right so i've seen data engineers who are brilliant i've seen data architects who are like extraordinary in my own career i've seen full stack developers who are brilliant i would not suggest them to move to machine learning i would say you become better at your own game because you're already so damn good become better at it like when i was at amazon some engineers in my team used to say hey verma i want to become machine learning engineer from a software development engineer i would ask them only once one very simple question you can i'm not denying that but you have two paths now from a software engineer you can become a senior software engineer or you can become a machine learning engineer if you're really passionate about math right take the machine learning engineer path if you think i don't have time i don't want to put in this much effort of time then go through the senior engineer path there's nothing wrong with it so there is great opportunity even in sales marketing business development there are phenomenal people that i come across every day hr finance so whatever you are you can excel in the same role no doubt about it but going into machine learning as an emerging field there are more roles there is more opportunity to grow and you will have the early more advantage cool so okay uh uh tahir has a question for a non-technical guy what sort of role is best very i wanted to actually answer that question i forgot in my notes so for somebody coming from a non-technical role it is easiest to crack the data analyst rule okay it is going to be the easiest and i we have we have had non-technical folks who also crack data scientist roles pure machine learning roles are applied scientist roles are less probable i'm not saying impossible less probable a more realistic thing that you can look for is join as a data analyst work in a data analyst role because it's easier to crack because the hiring bar is also slightly lower first crack that rule excel in the troll and then jump to a data scientist room that's what i would recommend especially if you're coming from a non-tech background a data analyst role is the most apt role for you to get into again we got some emails about students asking hey i have a career gap after my b tech what should i do if you have a career gap again always justifying career gaps are hard we have recommended this multiple times in multiple live sessions of ours don't quit your job to learn machine learning if you cannot balance your job in machine learning just leave it continue in your job when you get time then prepare machine learning because a career grab is often very very hard to justify we recommend everyone not to take that gap but unfortunately imagine if you have a gap first thing get into an internship role also if you can get into some role to show again we have had students with two years three years career gaps what we personally recommend to them is get into an internship role first very often what we have seen is these students are hungry and they are hard working they get into an internship role within three to six months they get a full-time role or they jump to a company if their company doesn't offer them a full-time role they use that three to six months experience and jump into a full-time role so if you have a career gap if you get a full-time role great if you don't get a full-time role it's perfectly okay to join in an intern role and then jump full time that because for internship role the company is also less critical it will give you an internship easily than a full-time role cool uh okay so uh aditya uh okay so let me okay aditya patak has a very nice question i want to answer that where to study machine learning and deep learning for free good textbooks i mean they're very good trust me if if i were to read learn machine learning on my own the best place i would do is from the textbooks that i was telling you right pick any good python programming from orelly pick any pick think stats which is freely available online i think they've opened it's from greenpeace publication so think stats to learn basics of statistics pick a pattern recognition christopher bishop or i mean this is slightly mathematically intricate but you you'll get a gist of it don't worry okay deep learning these standard textbooks are i think one of the best sources i know of because they have both mathematical rigor mathematical depth and if you just implement all these stuff from scratch on your own there is nothing that beats that again look at it this way right where do all the top universities do all the top universities use these textbooks as as textbooks and reference books in their courses they're not foolish right these are people who have been teaching who have been educating people for decades so that's one of the best that's one of the best sources that i know of okay mukesh says are there data science jobs in kolkata much fewer to be honest with you they're they're much fewer than in bangalore hyderabad puna or or delhi ncr right much fewer in kolkata there are some even in mumbai uh so lakshit says can we get applied scientist role as a fresher yes yes yes in product based companies yes we've had some students of ours who have cracked applied scientists role at the top fang like companies or equivalent companies as applied scientists yes the bar is just higher that's it and these are typically people who are very very strong in math fairly strong in math okay so uh rishabh very good question what do you mean by first principles when deriving algorithms very nice question i i am happy you asked that so when i mean first principles okay let me ask this question let's take a question from probably our high school days you have pythagoras theorem right everybody knows pythagoras theorem right you have a right angle triangle this square plus this square equals to this square right it's called base and height height square base square is hypotenuse square derive it we all did it in high school we all have forgotten how to do it when i say from first principles derive it using basic mathematical concepts that's what i mean again you can you can prove the pythagorean theorem again you can't use trigonometry to do that because trigonometry is based on pythagorean the pythagoras theorem to some extent right you can't use trigonometry to prove that okay so you can do it using simple geometric construction of squares and rectangles and things like that so that's what we learned probably in our eighth or ninth class if i'm not wrong similarly when i say derive these machine learning algorithms first from first principle you can say for linear regression okay what do i want to do i want to build a hyperplane given some data i have y's i have x i's i want to fit a hyperplane what is the equation of the hyperplane write that what does each of the pieces of the hyper plane mean okay if i want to fit a hyper plane what is optimization problem i have to solve why is this why is this loss the right metric how do i regularize this model so that's what i mean when i say build it from first principles okay so that that's the core idea here so uh okay space clones i don't know i think that's an alias he says he or she says can we complete interplay course in two months i think that's less probable no so i'm i mean we have had students who have completed in three months also but two months is unless you already know a lot of machine learning you already are very comfortable with mathematics and programming it's hard to do it in two months so i won't recommend that realistically six months is a good time for you to complete it right cool uh so uh kiran deep asks does freelancing in machine learning increases in upcoming years like web development could be there are already some of our folks who do some freelancing work some of our students also doing freelancing work but i think one problem in machine learning freelancing would be sharing of data companies are little reluctant to share their data publicly with a non-employee in web development they just have to give you the requirements doc and you design the whole website it's easier there but people are very very reluctant to share their large amounts of public that large amounts of data publicly with others so freelancing has a scope no doubt about it but i don't know to what extent the privacy concerns and data secrecy concerns of companies could play a role there cool so naman asks do you have plans for blockchain not enough not in 2021 cool amit has a very good question what expectation from a four to five year experienced software engineer for career transition first and foremost as a four to five year career trans as a software engineer they would expect you to be very good in programming fairly good in concepts like multi-processing multi-threading handle all the boundary cases data structures even sometimes a little bit of object-oriented programming right because you've been in this field for four or five years right so they expect you to be reasonable with object loaded programming also that's one number two the expectation on on the machine learning front is more or less the same except for the scientist role they expect you to know the underlying math uh again for for people with your sort of experience one very popular question here is they'll take an algorithm and say hey how would you implement it using multi-threading or multi-processing in python right so that's a very common question asked to people who are coming from four or five years of software engineering experience to machine learning because in that question the interviewer can understand your depth of understanding of the machine learning technique and your ability to apply your prior experience to machine learning your software engineering experience that's one in terms of probability statistics everything that i told you will be will be considered again depending on the company the complexity of problems the hardness of problems will increase as you go to top product based fine like companies obviously right but broadly speaking you would have probably more rounds of programming probably more rounds of data structures algorithms to some extent surely some of your questions will focus more on productionization of models because with four or five years of experience you would have spent time productionizing your code right so there will be some questions on that front also cool so aditya asks will the course content of diploma program similar to applied a course there will be there will be some overlap but they will not be exactly the same right because it's an academic program right so again all the syllabus of the diploma with uoh has to be approved by the by university of hyderabad's core committee on behind the behind the diploma and ml program so we have already worked with them most of the syllabus is more or less ready but there will be some edits as we go so it will not be exactly the same there will be some overlap there might be more academic rigor more academic rigor in the university program cool uh so very interesting harsh harsh is a bsc in math and pursuing mca you can surely pursue harsh the the diploma program in a and machine learning from university of hyderabad because by the time you finish or by the time you finish your mci probably by the time you're in your third year of mca uh you'll also finish the diploma and that will give you the career readiness so again since you already have a bsc in math i'm sure you're comfortable with some some some basic math for sure at undergraduate level bsc so it'll be it'll be it's a it's a it's a match made in heaven actually so math background currently doing masters in computer science like an mca pursuing the diploma in a machine learning is like a perfect match okay for data analytics project what can we do to acquire the data like simply downloading from kaggle allen very good question alain david has a very nice question so okay the simplest and the easiest way is download it from kaggle or some other source again that's trivial right to be honest with you the best way is if you can obtain the data yourself so i'll give you the levels of data acquisition download from kaggle is the simplest it's a joke to be honest with you okay second scrape the data or obtain the data by calling apis okay we have done a live session on that i think it's publicly available also we have shown you how to acquire data using being a bing search api using amazon products api you can get all this data using apis to pitch data that shows a little bit of programming maturity ability to use web api is publicly available second third scrape the data using using using again try to write as much code as possible you can use selenium you can use beautiful soup all of these things that's the next level third go out and fetch the data yourself for example i'll give an example one of our students he wanted to build a self-driving car type of system for indian roads so he took his smartphone he stuck it to his car windshield and he drove it on indian roads okay and he also put a camera to see to see how much his wheel is turning so we have one such case study with us data so he went ahead put physical camera your standard smartphone cameras collected the data went ahead built a simple cnn model nothing complex in all of his interviews in all of his interviews he joined in a very good role very senior role i think he's somebody with about six seven years experience so very interestingly all of his interviewer said hey this data acquisition that you've done is terrific right most of his interviews were about data acquisition not about deep learning as much because he has done something unique so in terms of data acquisition downloading is trivial nobody cares about it use publicly available apis to fetch it that's better scrape the data the next best fourth best is go out and find ways to fetch the data yourself in an ingenious manner that is unbeatable okay cool uh so adroop says i have applied for again applied routes for those of you who are seeing this live stream applied routes is just a parent website where all of again earlier we had we were putting each of our courses at one place so we thought let's put everything under one umbrella term called applied roots and the name itself is because we focus a lot on applied aspects in every course that we teach and roots because we want the foundations of of your learning to be more applied in nature more fundamental in nature more foundational nature that's that's the reason for the name applied roots applied roots is also our website the diploma program that you applied for is run by play course itself okay so uh one second so uh i just i mean i just scrolled it and missed the question i almost half read it [Music] okay i will move on okay that's okay uh huh so siddhu km he says i have two years experience with masters in aeronautical engineering currently during data science course how can i showcase my experience while searching for a data science job can you attach a list of companies to apply siddhu there are two three things directions that you can look at try to pick up data sets from kaggle from engineering mechanical engineering from machine failure there are some of those data sets or there are also some data sets from fluid dynamics because i'm sure you would know what fluid dynamics are and things like that right so there is a lot of recent research in the last year or so on how to apply deep learning systems to solve complex fluid dynamics system fluid fluid dynamics basically pde is your partial differentially complex pdes so you can do your case studies on that also do a case study at least one of them in nlp the companies that you could look for are companies like boeing companies like abb right uh yeah abb bosch right all these are companies which are in that electrical engineering mechanical engineering maybe some aeronautics right so it is in that intersection that you can look at like the companies that come at the top of a honeywell yeah that's another company which works very closely in aeronautics and other air so companies like aeronaut i know abb works very much in abb is this huge i think uh swiss giant if i'm not wrong which manufactures high-end uh mechanical electro-mechanical equipment so these are the companies that you can look at for sure of course there are also other local companies when i say local these are also giants like tata motors again the auto industry is slightly in a slump now but those are like mahindra mahindra right all these companies are also setting up their data science machine learning teams that you can look at but don't limit yourself only to these companies do some case studies in nlp try to get into general data science also don't don't fixate yourself just on aeronautics or electrical or mechanical engineering it's good to be prepared for all sorts of roles if you get something in mechanical engineering or aeronautics that's great but try to be more broad in your search so although asks what is the scope of opportunities in gulf countries again i'm not very good at that because i don't have deep understanding of the market but at least one country that i know of which is investing heavily into ai is uh is arab emirates united arab uae dubai basically as many of us know so that's one country which is investing heavily into a machine learning and equivalent systems so again i don't know the job market very well there to be honest with you but i think that's one place where you should seriously look for okay uh oh very interesting adhurb has a question i'm working in infrastructure and code for four years i have knowledge on linux and windows servers i have red hat certified admin devops okay very interesting so you're a very ideal candidate for a career like ml ops now what is ml ops ml ops is the whole devops thing but for machine learning right in devops what do we do we take care of the life cycle of code right in machine learning as in ml ops we take care of the life cycle of machine learning code and machine learning models so for ml ops your skill set your four years experience is very much aligned to it you just have to understand machine learning understand how how the understand the whole life cycle understand what each algorithms do so even if your mathematical understanding is not too deep you're still okay because devops like sorry ml ops requires you to have a like even if you have a data analyst like knowledge right you know you know everything at a high level you may not be very good into the deep and deep mathematics that's okay if you understand the applied aspects if you understand how to apply where to apply what happens after this what are the problems that you will encounter here if you know all of those intuitive details and if you can combine that with your four years experience in in in as a devops or as a system admin i think that's a nice that's a nice role that you can get into cool uh ah where coding skills is more required sda role versus data scientist role surely software development engineer for a software development engineer the coding skills required are surely more than data scientists no doubts about it hands down okay so purushottam has a question so this i'm i'm happy you asked the question pursuatham he says what is the importance of knowing big data spark all these things for machine learning roles again these are good to know skills these are not mandatory okay if you know them great so in our course we have covered about nosql databases we have covered we have done multiple sessions on spark both at a system level as well as how to up how to use spark for machine learning applications and data science applications so spark is if you know it great if you don't know it it's not it's not a road blocker because if you're strong with foundational concepts if you know programming picking up spark is not very hard and again most data scientists when they use spark they use spark sql which is very similar to your uh standard sql so if you know sql picking up spark sql is not very hard and very often you may not you may not own the spark cluster you may not be responsible for running the spark cluster because very often you will have operations engineers or you'll have software engineers who are managing that for you so knowing it doesn't hurt it's always good to know but not knowing it is also okay it's it's not a mandatory skill it's a good to have skill cool so i am 2019 pass out so uh i don't know i i think this is an alias so he says i am 2019 pass out still not got a job is it possible to get ml jobs so my suggestion for you is this i'm sure 2020 has been a hard year for you because the whole job market took a nosedive because of lockdowns and things like that my suggestion for you is instead of increasing your career gap try and first get into an internship based on what you currently know that's what i would recommend because learning machine learning transitioning will take time why even waste the time if you already know programming let's say get into a programming internship or a full-time role first and then thoroughly learn machine learning that's what we would personally recommend okay where is this so psy asks i'm doing my m tech in ocean engineering from iit madras can i switch to ml by doing my thesis in ml that's certainly a good idea sai so i we i have encountered people from non-cs backgrounds but i'm sure the ocean engineering department is very heavily math driven so if since you come from a top university if you can build if you if you can know if you can become comfortable with programming understand the math which i'm sure because you would have understood lot of complex math underlying ocean engineering learning machine learning shouldn't be that hard right so please pick it up and as part of your master's thesis if you can apply what you're doing what you've learned in machine learning and deep learning to ocean engineering that's great it will surely open up doors in machine learning for you but the skills don't change whether you come from an iit or you come from a staff you come from stanford the bar only increases i've seen this including myself when i go to a top university i tend to ask harder questions because i expect them to know it for the same role for the same role when i go to bed when i go to a more top university like indians of science iit stanford etc i tend to ask harder questions because i expect these folks to be better in general right so the expectation from a top university graduate is also higher so you have to be you your comfort with math should be really really good again because you come from a non-cs background they can't test you much on programming so they would rather test you more on your math stuff that's where you have to be really good at cool so sriram has a question is what are your thoughts on the bsc online degree from iit madras again i'm not sure how it is again i've heard about it i know how it's being operated we'll have to wait and see how this pans out it's too early for me to comment on it i mean it's it's a nice idea no doubt about it but how does it work out we'll have to wait and see okay this is a good question gareth should i join a startup or a big mnc in data science or machine learning as a fresher i am from mechanical engineering background there is nothing wrong with any of them but my recommendation always is don't pick a company pick a team whether you join a large company or a small startup ask yourself especially early in your career right it is important that you have challenging work that's very very important i can't like at amazon when i joined i was one of the earliest machine learning scientists in my whole organization that gave me enormous amount of learning experience when i joined yahoo labs fresh out of college it was a very mature organization with phenomenal engineers and scientists i did not get much to learn so whether it is a large company or a small startup try and pick a team where you will get to do things in the next one year you will get to build things from scratch if possible because that will be enormous amount of learning i can't say i mean that's that's better than joining a well-established team because you may not get to do great work there so always my my i would favor a team which is early uh which is which has a clear direction for the next one year where i would get to learn a lot irrespective whether it's a startup or a big company it is more likely to find such teams in startups than in bigger companies you have to get lucky in a bigger company to find a team like that in smaller companies there is a higher chance you find such a team and hence again it's it's a pick and choose assuming the compensation is same i would pick at least early in my career even later in my career actually i'm a type of guy who want challenging work so i would pick that in a day cool yes shubham is asking can you please provide the links i will provide the whole the the pdf that i discussed i will post it in the in the description section in a while okay very good so aditya has a question so where is that question we put all the kadamaka stuff okay so he's saying i'm i'm i'm writing my gate exam now so which is better going to indians of science or iit doing my masters there or pursuing a pled a course my suggestion here is going to a top university will open up lots of doors for you let's be realistic there okay so my suggestion for you is as follows why don't you crack gate now you're in your third year right crack your gate now in your final year prepare much prepare learn ai machine learning try to get a job again this is something that we told all of our gate students also right after btec if you're getting into amazon or google or flipkart or facebook or any of these companies right or microsoft don't bother going into going for a master's program unless you want to become a professor you want to do a phd i mean there's no point in it because at the end of it you'll most if job is your goal you already are getting the job that you would get after iit or isc right realistically speaking so my suggestion to you is uh prepare for gate it will make your foundations very strong it will make you a good computer science engineer crack gate if you crack gate great if not you can always pick up the ai course after you graduate after you go to a job or even we have a lot of students from iisc and iits who pursue our courses because they want to have a leg up in competition in their job search or in their job interviews so either of it is all right so arvin says what are applied a plans on ai topics in 2021 we are trying to cover more cutting edge topics in our live sessions we're planning to conduct more interview sessions in our live sessions uh for example uh this is something that our students like a lot and it's also beneficial to the students so we would pick up each chapter or each module and we would have an interactive session where we will solve problems together where i'll ask pointers from students and things like that right now we are doing this interview session type of setup for model productionization like last week we picked a couple of case studies of ours and we discussed pros and cons of each each productionization option because there's no one app productionization option right if you if you suggest me something i can give you a counter example where that won't work and i'll keep asking you follow-up questions right so i think we'll focus more and more on this interactive interview question sessions while covering more advanced concepts as they evolve so one thing i mean there are a bunch of concepts at the top of my mind so surely we'll cover all this again pytorch is something that probably will cover very soon so there are some of these again we may not be able to go full deep into pythons but we'll give an introduction so that you can pick it up and run with it cool uh very good so javid says i have three years experience as data scientist in startups and small size companies what should be my approach to get into big mnc and top product based companies very simple javit you become better in programming become better at math because given your startup experience if you apply to a if you apply to some of these companies you have a decent chance of being being asked to come for a screening test or something like that you have to prepare well for it because most companies for data science roles they have at least amazon like companies they have a screening test then they have about five six interviews so first prepare thoroughly and then start applying for these companies it is likely for you for you to get a call because you already have three years experience at interesting startups so when you get a call be prepared to crack it that's all i would recommend okay uh uh so uh one second uh so for ml ops what is the so this is from uh okay this is one of the students which is from ml ops what is the approach should i follow can i can you make a session on it sure i will probably will we'll do a session on ml ops i mean that or i'll i'll put out a video i probably will do a live session on mlaps i'll do that for sure no problem uh so what level of coding is needed for a data scientist very nice question vikram that depends on the company so let me give you some examples right um take amazon okay take amazon or fang the top product based companies right they would expect you to be able to implement an algorithm with reasonable space and time complexity using inbuilt data structures they are not expecting you to be an expert in data structures algorithms for example they might ask you how do you implement pandas dot merge which is like a sql join operation right they'll give you two tables and say merge these two tables write code from scratch or they might ask you something like how do you implement x power y examples that i gave you right so or they might ask you hey i'm giving you some hundred numbers find the kth smallest number how efficiently can you find it so there will be some focus on time and space complexity at top product based companies one level below it'll be easier like how do you slightly easier questions but even these questions are not hard if you know how to write for loops if you know the basic inbuilt data structures like hash tables or dictionaries in python lists etc this is fairly straightforward these are some of the harder questions that you can expect or some of the questions you can expect for data scientist tools at the top companies and some companies might just give you a task for data scientist roles they'll give you a data set and say hey do all the analysis using any library of your choice that's more of a hands-on thing so then you say hey plot pdf cdfs you have to know matplotlib all the standard libraries that's another type of interview that some companies conduct and then what they do is they give you two hours they give you a data set they'll make you analyze the whole thing and the interview will be hey why did you do this why did you do that what is how do you interpret this plot what do you make out of this and things like that okay so this question is about tiny ml again do you see openings in india not much to be honest with you because we as a country are not as much into hardware development if if if if you would ask me about china i would say yes there are options um not so much in india to be honest with you that's why it has never been the focus of our at least the mark job market that we see there are very few companies who say hey we want somebody who is an expert in machine learning for embedded processors we don't see that much because what happens very often here is uh it's typically an embedded systems engineer or an embedded system software engineer will work with machine learning and they both will build it because it's very very rare to find the talent and the number of openings are also fewer i at least at least for the next year or so i don't see that taking off to be a big role cool so uh yash has a good question he says i'm a qa automation engineer transitioning to data science profiles will my experience help yes so for example as a qa automation engineer i'm sure you know things like selenium right or qtp so you can pick up selenium selenium you can use selenium with python by the way so why don't you i mean in one of your case studies you want to build on top of your experience in one of your case studies use selenium or any of these equivalent tools obtain data from the internet right by scraping it carefully of course respecting robots.txt fetch the data build some machine learning models on top of that now what are you doing you're showcasing that you know machine learning you also are building on top of your prior experience shortly there are i mean you can surely transition from a qa engineer type of role because as a qa automation engineer i'm sure you would write some code so it's possible even for people who don't code to transition to data science roles but for you it'll be easier okay cool uh so during the interview process at product based companies are we allowed to use google no no no most often no there might be some screening tests where they might allow you to use google but not during the interview process suppose if i'm asking you something about let's say gbdt's or something about convolutional neural networks if you say hey i'll google it and find it i'm like sorry but let's say you might give you a two-hour task again they'll tell you it's an open book task suppose you've forgotten a function you can always google search for that function and put it back that i think most people are okay with okay i'm sorry just a second folks so niraj has a good question are the roles in the industry less challenging than research how to stay updated without doing a phd for example i don't have a phd i chose not to do it i mean because i wanted to be in the industry learn how to get things done so the way i keep myself up to date because i also interview phd students i've interviewed i've hired phd students i've interviewed them so i have to be at the top of my game obviously so a simple rule of thumb that i would suggest anybody who wants to stay at the top of their game is at least allocate five hours a week on self learning pick up the best research papers my simple strategies on twitter i follow this bunch of phenomenal researchers in ai machine learning i follow google research i mean google research deepmind facebook ai research amazon.science i follow a bunch of these labs even university research labs i follow university research labs i follow top professors top researchers in these fields when they talk about a paper i try to pick one paper a week and read it sometimes it might take me two weeks to finish a paper if it's a slightly harder paper but i try to do one paper a week that gives me 50 papers in a year wow that's that's terrific 50 papers in a in a year if you do it consistently like i do almost i mean i think this year i'll have to count it i think i did something like 35 odd papers new papers this year completely new papers because i was also reading other stuff uh on how to teach better uh how to i mean i was doing something on education research and things like that but i think i did roughly around 33 to 35 papers this year so i'm almost on the top of my game because i know i know exactly what's happening because i read all the top papers from the most recent conferences so if you stop doing that you become a dinosaur in just a few years so the reason again is something that i recommend recommend to everyone who is already a data scientist okay okay folks and please understand that it's impossible to answer every question in a public live session it's almost impossible realistically so i've tried my best and all only my throat is hurting because i'm speaking non-stop for two hours now so please email us any questions that we could not answer here at team at applied ai course we'll try and answer them as fast as possible okay sounds good thank you very much and have a safe year end again please be extra cautious given given a surge in cases in some parts of india and the world so please be extra cautious please stay safe let's wait for a few months till the time all of our all of us are vaccinated and i hope all of you have a good year ahead i hope most of you they're like already 600 plus who are watching live i hope all of you make a commitment and promise to yourself that next year you will spend these 12 15 hours on self improvement it could be machine learning it could be anything if you want a career transition try to at least spend 10 to 15 hours every week on it if you want to excel in your current career try to spend five to six hours if we do that i think the you will be at a much better place and surely the world also will be at a much better place okay thank you very much and advance happy new year to all of you and a belated christmas to everyone who is celebrating i mean we have a big christmas tree because my son is crazy about christmas trees so yeah thank you very much folks have a good year please stay safe i can't emphasize this and thanks to everyone um who i mean this was a rough year for everyone including us there's a lot of new stuff that happened for us it has been a rough year i completely understand that i'm sure it has been a rough year for some of you i want to thank all of our students all of our well-wishers everyone from the bottom of my heart and on behalf of all of our team who have stayed with us in this year in this very rough year in terms of lot of things and again let's hope the next year is better for all of us thanks everyone who has attended this session and most importantly who helped us throughout this year in coming out of some rough patches thank you very much folks thank you

Channel: Applied AI Course

Views: 36,709

Rating: 4.9421277 out of 5

Keywords:

Id: 7CH8Mb6OH3k

Length: 124min 20sec (7460 seconds)

Published: Sun Dec 27 2020