How to prepare for Machine Learning interviews- Part 1 | Applied AI Course

hey folks uh thank you for joining in let me just check if everything is working alright just give me a second I just want to see if everything is working all right if folks thank you for joining in let me just check if everything is working on yeah so so this yeah I can hear myself in the live feed so I'm assuming everything is working for everyone I'm also keeping a I'm also like looking out at the chat window so that I can trick as many questions as we can today's session so yeah so let's get going or since since it's exactly 10:00 let's wait for a couple of minutes I'll answer a few questions hey thank you folks hi hi everyone thank you for joining in and let's just wait on the on the live chat here and let's wait for everyone to join in there are a couple of points that I noted um that I thought I will cover and these are mostly about general strategies for interviews for freshers for experienced folks folks with different types of experience folks from different engineering backgrounds different disciplines again different experience levels different companies different roles etc I've made some notes I'll try to cover as much of this as possible in this live session but I'll also I'll also try and shed some light on what depth because I see this question asked where many people should remember formulae certainly why not so again depends on the company you're interviewing for depending on the team that you're interviewing for what more people are what what most people are looking for is not whether you remember a formula by heart I don't care if you remember Bayes theorem by heart or not but do you know what Bayes theorem actually is if it they're a lot of formulae that I do not remember for example if you suddenly ask me okay what is the primal and even of his film I don't remember it by heart but if you give me two to five minutes I can actually derive it okay similarly if ask me any any formula like that right of course I know some basic formula like what is entropy these are some basic stuff that I do know but if you ask me to derive if you ask me some slightly complex formula I may not remember it by heart but I know the concept behind it that I can derive it in like two to five minutes and that's what is important actually I have it I'm terrible at remembering formulae but formulas are important because they tell you how things behave if you don't know the Bayes theorem right I can't expect you to know how big have naive Bayes works right without knowing Bayes theorem how can I assume that you understand nine base if you can't explain me what is the loss function of logistic regression you can't you can't understand how logistic regression behaves when you have outliers or when how logistic regression behaves when there are some some anomalies in the data how it behaves when you have imbalanced datasets all of that can be derived very easily and very ingeniously from the formulae so actually just like many of you I'm terrible at remembering formulae but I can derive formula from first principles and that's what we have done in the course also right if you realize in the course videos except for a very few formulae most of them we have tried to derive from very first very basic principles in geometry and algebra okay your your 12th class mathematics is all you need to derive most of machine learning most of machine learning formulae actually of course there are some additional stuff that you learn but we've discussed most of them in the course so instead of trying to remember formulae remember how they are arrived at and the intuition behind it in the geometric or mathematical right that's super duper important so so let let me so I'll again to be honest with you I'll try to focus this discussion only on how to prepare for interviews I try to avoid taking questions from the whole spectrum of topics because that way I will not be able to cover the topics that I want to cover and address questions from students who are here especially from that topic so instead of trying to answer questions across the whole breadth ins a whole breadth and length of machine learning deep learning I'll try to narrowly focus on how to prepare for interviews based on your background okay so I see a lot of questions here and since we waited for a few minutes I think none of you have joined thanks for joining and a very good morning to everyone who has joined in so let me get started anyway keep a lookout for all the for all the for all the comments in the chat session and before we get so the focus of today's discussion again I've just written some notes here and I'll try to discuss as many points as possible so first and foremost okay first let me give you an overview I'll also answer questions as we go forward right first let me give you an overview because in this overview almost 80 to 90% of equations will be covered that some of you are asking me that you have computer science experience of non computer science experience how do you how can you transition to machine learning roles I'll answer all of that so let me give you a broad overview I'll keep taking questions in between I'll try to make it interesting they're also very good question on if I'm an MBA how do I transition to machine learning roles I have that covered in my notes right so let's go step by step first and foremost here I'm assuming that you know some machine learning you know some basic programming before you go to much learning interview right so we'll focus on how I'm assuming that you know some machine learning deep learning some mathematics some programming and things like that without that going for a machine learning interview is useless it's like going to a software engineering interview without knowing what programming is it's useless so don't even don't even apply to machine learning roles without knowing decent amount of machine learning and what is decent we'll discuss that through the rest of this session so first and foremost interviews come in all shapes and sizes I have seen interviews in the hundreds of interviews that I've done just at times in itself I have interviews come in tons of shape and the hundreds of students that we have placed as part of applied a a course based on the feedback that the students have given in in almost 300 plus companies snuff we have seen a huge spectrum of interviews and I mean there is a huge variety right but we can break that variety into some small sections I'm not saying that I can cover every type of interview right because there are as many interviews as there are interviewers and very importantly since data science and machine learning is a more recent type of job unlike software engineers because software engineer as a job has been there for the last 30 years right since the early 90s this job of a software engineer has been there so software engineer interviews are much more structured because there is a structure that is associated with for example for software engineer interviews if you're interviewing for a at a services company especially as a fresher you are typically ask basic questions in programming your to write some simple programming questions like implement a queue implement a stack write or you are asked basically if it's a services company typically are you given a simple set of problems which requires some basic data structures basic algorithms and you can implement it that's one type of questions that you are asked that enforces TCS all these companies also you might be asked some basics of databases like SQL queries again this is all very basic stuff that will be asked again but if you go to a product based company like Amazon or Google the level of questions will increase but typically most of the questions are again from data structures algorithms some companies like some some companies also may have aptitude rounds depending on the company so there is a much more neat structure around software engineer interviews but for machine learning and data science it's not still very structured ok it's still there still there is still a lot of gaps in the there is no structure because this role is also still evolving but over the last decade that I've been interviewing candidates I have seen some structure evolve ok it's not perfect again I'm not saying that every interview you may attend in your life will fall into the structure but there is a there is some structure here so we'll try and cover the most type of interviews here so let's let's go here ok typically typically whether you're fresher or an experienced candidate with very again first we will see what are typical rounds you'll encounter right typically 99% you will have a programming round again the complexity of the programming round will depend on the company will depend on the team will depend on the will depend on your background ok but broadly speaking you will have a programming round of some sort obviously this programming round may test anything from basic programming skills or your SQL especially because data scientists and machine learning engineers end up using lot of data sources like data bases data warehouses etc SQL is something that is often asked if you say I don't know SQL I've never studied SQL I'll say that's ok again it depends on your background suppose if you are a mathematics PhD or a physics PhD right or if you are a civil engineer who is never studied or if you are mechanically who is never studied SQL SQL is not expected as much out of you but if you're a computer science engineer or you a working professional who work in software engineering for a few years then you are expected to know SQL right so programming and SQL also again we'll see what types of questions are asked based on the teams and based on the company's in a little while but typically there is one or two rounds of programming and of course if you go to a services company the level of questions is certainly easier they might ask you simple questions like ok let's take a few questions right and very simple questions you are given two sorted lists right or two sorted arrays and you define the median of these two errors medium not the mean the medium of these two errors which is again very simple problem because these errors are already sorted this required this is literally if you write code in Python this is literally ten lines of code right so these are very simple problems where they're just testing if you can program in a programming language using loops that's all they're trying to test they are not trying to test whether you know some advanced data structures or some advanced algorithms time complexity space complexity all of that okay they are not testing if you know KD trees all these advanced data structures they're asking some very very basic stuff okay again the first maybe you could have one or two rounds of programming again the number of programming rounds also depends on the team ok first let me give you an overview of all the types of interviews that you could encounter one is programming programming could have programming some data structures and algorithms again depending on the company you can also have SQL in your interviews ok we have covered all the important topics as part of our course anyway I'll come to what we have covered in the course little later the second type of thing which is very very important is called real-world problem solving rounds these are these are some of the best interviews actually and that's because there is a reason behind it and that's because you might learn all the theory Under the Sun but if you cannot solve a real-world problem what's the use of all of that and these and these problems are more focused on experienced folks suppose if you are a software engineering with software engineer with four years experience you would have learned the ability to solve real world problems in these four years right so tea then you're asked the problem like this okay here is here is a real-world problem for example they may say okay we are a music company okay let's say Yuma you're working with Saban okay which is a which is a music streaming company in India so or or let's take Spotify or any of these companies like music streaming companies like Spotify or so on they might say how do you recommend new songs to people very simple question very real-world question then they expect you to work out the solution from what type of data you collect - how will you productionize the model so these interviews typically are very open-ended they'll give you a problem a real-world problem most likely interviewers give you problems that they are working on themselves because they understand the problem very well right so I have a half and asked in my interviews people have explained them a small module of the problem that I am solving currently and asked them how do you tackle this because as an interviewer this gives me a lot of insight into how you think how you will be as a colleague if I hire you these are these are extremely extremely valuable interviews especially for experienced folks if you're a fresher you could give some textbook solution and people may say ok this person does not have a lot of experience so that's ok but even if you are an experienced software engineer people expect you to have that acumen to have that insight and how real-world problems are solved in the industry and be able to translate that right so real-world problem solving interviews at superb and that's actually we often ask our students to build a strong portfolio of at least two projects preferably up to five projects that's because when you have a portfolio of projects and when you take it to the interviewer the interviewer tends to focus a lot of their time on the problems you have solved because if you cannot explain me the problem that you claim that you have solved thoroughly then what will you solve a new problem so that's why portfolio of projects is super duper important we emphasize so much on it as part of our course and in general also you should always solve two to five problems thoroughly end-to-end you should try to solve these problems in doing everything from obtaining data cleaning data pre-processing data building the whole models coming up with the right matrix to production icing these models that whole spectrum you should be able to solve at least two to five problems based on your domain for example somebody is asking here that he comes from bioinformatics background if you're if you're coming from bioinformatics background then obviously you should solve by dramatics problems the tons of bioinformatics problems which use machine learning and data science if you come from banking IT domain and suppose imagine you worked at a bank as a full-time employee or as a consultant then you need to understand I you hi folks our I'm back let me just check if everything is working sorry I had a power cut extremely sorry for that and yeah sorry somebody says I I'm on breakfast it's been a long time that I had a breakfast I typically wake up at 4:35 so I finished my video I finished my breakfast by 6:00 a.m. anyway so back to our discussion sorry I had a power cut the summer is still not over in Hyderabad and it takes a few minutes for my inverter and my Wi-Fi to get started again sorry folks sorry for that anyway back to the discussion I hope everybody is back here right it's working right for everyone just refresh this ski just refresh the screen and everything should work because I just joined back in immediately reconnected me back power cut prediction where do you want to do that you can just ask the eliciting Department and they'll tell you when the power cut is anyway sorry sorry sorry for this goof up yeah I hope it's working fine yeah so where were we um we were discussing about the real-world problem-solving part right so unfortunately the power gets cut again I'll just join on my 4G in a few minutes but typically what happens at my residence is met the power the the inverter kicks in within a few minutes right so just stay online for about a couple of minutes and we'll be back there ok so back to the discussion so we were talking about real-world problem-solving questions and these are some of my favorite questions because they helped me understand as an interviewer how you will be as a candidate that's very important for me right because it will help me it helped me understand how you will be as not just a candidate but as a colleague if you join my team and if we are working on the problem together how does your thought process how do how can you think that your real-world problem-solving is SuperDuper important wrong that I think almost every company across the spectrum from the smallest of companies to the world's best companies they typically end up having one of these real world problems and these real world problems can be picked in two ways number one the interviewer could choose the problem that they are solving currently and give it to you as an open-ended problem to see how you think number one now - is they could also ask you a problem that you have solved in the most recent past again I've seen this in hundreds of interviews right if you try to fool the interviewer if the interviewer is even half smart we can detect it I've seen candidates who claim that they have done this they have done that I said okay great let's dive into it so I start asking some of the questions that only a person who has actually solved it can answer and they get caught we immediately come to the call and say thank you we're not interested in you right so please don't try to cheat an interviewer because if he's even half smart you can get caught only if you have solved a problem really only then explain that problem to the interviewer that's very very important because interviewers especially people have tons of experience interviewing will catch you in like no time I've done this I've seen I've seen lot of candidates who claim they built that they've built that but when you go slightly deep where real world stuff matters they'll get caught so please don't Bluff an interviewer but most importantly solve a portfolio of projects at least two projects end to end so because that gives the interviewer a chance to dive into problems that you solved again in in an interview as much control as the interviewer has there is equal amount actually I would say there is more amount of control of the interview by the kinetic that a candidate can control the interview much more than an interviewer for example I've done this many times I've seen lot of smart candidates control the interview they take the interview in the direction that they are comfortable in instead of the direction that the interviewer is comfortable that way they can actually showcase their strengths right you don't want to be stuck in an area of machine learning or or on a problem that you're not comfortable in because that way you're not able to showcase your skills so a good candidate knows how to take the interview how to guide the interview in the direction that he can perform well in that's very very important right so if you have a portfolio of projects the portfolio of projects are like are like good catch points for the interviewer to pick on for example if you if you come to me for an interview and if you say I've solved this problem in recommender systems I'm very good okay you've solved this so let's dive into recommender system stuff because that gives me a hook into one topic which I can dive into deeply so it's very very important you have a portfolio of projects for the real world problem solving around then so this is a second one first is programing second is real world problem solving rounds the third one the third one is machine learning breath machine learning Bret Thrones again many people may not use this term called machine learning breath but typically there are rounds wherein they try to test a wide spectrum of algorithms that you know okay they try to test if you know let's say if you know this 10 or 20 most widely-used algorithms or not so those are called as breadcrumbs why are they called bread because they are trying to test or understand the breadth of knowledge you have and here again every person has a different breadth of knowledge it is impossible for any single person to have the whole breadth of topics unless you've been doing machine learning for 2 to 3 decades to be honest with you there are some areas of machine learning where I am NOT super back ok where I am not a fact frankly speaking that I'm not an expert at it because I've not worked in those areas or after graduate school I've not worked I've not studied them immensely right so the breadth of knowledge gives the interviewer or knowledge so in a breath around rate they might ask you one question about one of the major techniques and in an interview if you don't know a technique for example somebody asks you about let's pick somebody asks you about conditional random fields let's say let's say okay so if somebody which which is a topic in which is a topic in graphical models in machine learning if somebody asks about conditional random fields okay you can quietly say I have heard of conditional random fields that this is what they do but I don't know exactly in depth what the how their use or I know or you can say I know where they are used but I don't know the exact mathematics behind it it is perfectly alright to say I don't know CRFs it's perfectly all right because interviewers generally don't expect you to know everything and because the interviewer himself most likely doesn't know everything right this is true with me as an interview I've done hundreds of interviews there are some topics in machine learning and deep learning that I am not good at and if an interviewer say gives me a solution based on the I'll clearly tell him sorry dude I am NOT an expert in this area can we try something else and that honesty really pays off trust me right so in breath around try and answer they'll try to test your breadth of knowledge they'll go from simple algorithms like naive Bayes k nearest neighbor logistic integration SVM's random forest GBD tease the whole spectrum of techniques they might pick one one small small small small question from each of them to see if you know all of the most important techniques or not and if you don't know something it's perfectly okay to say so because they are only trying to test or trying to understand what you know okay so those are called machine learning net rounds then very importantly there is something what is a machine learning depth drop these are very important because you may not know every technique okay that's perfectly valid but in machine learning depth round you are asked to pick a few techniques that you're very comfortable with or that you have used in the recent past okay for example let's say you might have used SVM's in the last one year in some real-world project there I'm expected to know SVM's very well because I've actually used it to solve a problem right in my as part of my portfolio of projects or as part of the experiments that I've been doing so when they ask you tell me a technique this is something that I've done a lot right in my own interviews and see lot of people also do this as interviewers they'll ask you to pick a technique that you're most comfortable with or they'll ask questions from the technique that you claimed that you have used in your portfolio of projects or in your most recent work suppose if I've used SVM's and say okay let's dive into SVM's I'll ask you every nook and corner of SVM every nook and corner I'll ask you right because if you have actually used SVM I assume that you have understood it thoroughly right again you need not be an expert in all the techniques but those techniques that you claim that you have used recently or the techniques that you claim you're good at you better be good at them because that tells me whether you are thorough with learning a concept and using the concept of not if you tell me that you're good at SVM but if you can't write me the hinge loss interpretation of SEO and I am like come on dude I mean this is not this is not accepted or if you can't tell me how support vectors are used right or if you can't tell me if you can't write down the to ization problem of SVM right again you just telling me the theory is not sufficient you have to explain me how kernelization is useful how do you design the right Colonel how do you tune the parameters of the right Colonel these are all important questions if you say that I have actually used these films or I am good at s films I expect you to answer these questions right again as part of our course videos we try to cover as much as possible again nor in its no single course in the universe can cover everything just like no single textbook can cover everything right so we try to cover as much as possible as part of our coursework again we give additional links also so depth round you have to be very careful and my suggestion is if you're building a portfolio of projects those techniques that you've used in your in your case study or in your projects of portfolio of projects try to dive deep into them and understand every nook and corner okay because that because the the depth rounds tell me how deep is your understanding of a subset of concepts okay then very importantly so these are these are the four types of interviews and there is a fifth type of interview this is your mathematics foundations lot of people mess up this interview people say okay I know again I've seen again we have done interviews and some of the world's best universities I mean some of my friends were also our mentors and advisors at applied any course have done I mean between all the advisers and everybody would have done thousands of interviews right we have seen candidates who can tell everything about every algorithm but if you ask them something in basic linear algebra or something in basic probability like we give them a simple problem in Bayes theorem we expect them to solve it you just have to apply Bayes theorem and solve that problem some people can do that right that shows that your understanding is superficial because your mathematical foundations are weaker and why is this from the important for an interviewer because new techniques will keep evolving there is no static Ness in machine learning and deep learning every week every month there are a new set of techniques that come right so that that is a problem right so if you if you want to if you do not want a sorry if you do not want to fall back if you don't want to fail in that right so the most in important aspect here is format found for again the interviewer is trying to test whether you can cope up with new techniques that evolve for that your foundational mathematics is important can you write the equation of a line can you compute the distance of a point from a line without that how do you understand well just take a digression how do you understand support vector machines right can you understand what is entropy right when the when does something have maximum entropy or minimum entropy these are basic foundations right without that you can't build a random forest or even a decision tree similarly a basic minimum Maxima of stuff basic calculus how does derivative change when you reach a minimum right there are very basic stuff if you can't ask answer these very very basic things mathematical foundations right these are typically your matrix algebra or linear algebra your probability Basics in calculus these are the three most important topics for machine learning probability and statistics linear algebra and calculus again calculus you don't have to be an expert in integrating using tons of methods the most important calculus for machine learning is differentiation makes again you don't have to be an expert in second order differential equations all of that the calculus that is required which is the maximum minima computation which is used in stochastic gradient descent gradient descent algorithms that part you have to be pretty good at you don't have to be an expert in second order differential those are not required because we don't use them day to day in machine learning of course if you know second-order differential equations that's great right but it's not important on a day to day basis right so that's very important again somebody is asking here that again most interviewers some there I've seen some interviewers also this is based on the feedback that we got from our students who attended interviews and who got placed some interviews just ask you a formula right they expect you to remember a formula just be be truthful to them and say I don't remember the formula but if you give me two to three minutes I can derive it for you I actually actually that that's that's what is important right or I can explain you how this formula works right and that that is very very important remembering by heart is useless trust me because what you remember today one year down the line you will forget but if you know how things are I'm net that skill stays with you for a lifetime right that's why that's why as part of our course we try to derive everything from from Haas principles from from what is the equation of a line what is the distance of a point to a line right and then slowly work out slowly work over then we give justifications and why we are doing this why we are not doing it right so that is very very important again somebody saying is learning unsupervised learning important why not see if you go for a machine learning interview the basic set of techniques you have to know if you say I don't know what K means is come on that's a joke okay if you don't know what is agglomerative clustering or or basically hierarchical clustering that's a joke it's it's considered one of the basic concepts right if you say I don't know some advanced some advanced stick suppose if you say I don't know DB scam okay that's okay because there are some some good machine learning engineers or some good folks who may not know DB scam but they can pick it up but if you don't know basic techniques that's a red flag for me if you don't know some advanced concept if you don't know something you

Channel: Applied AI Course

Views: 26,234

Rating: 4.9016895 out of 5

Keywords: #hangoutsonair, Hangouts On Air, #hoa, Data Science, Machine Learning, Deep Learning, NLP, AI, Big Data, Machine Learning Interviews

Id: sLAnpLlkh9U

Length: 34min 6sec (2046 seconds)

Published: Sat Jun 08 2019