Generative AI Interview Questions | iNeuron

that's for for okay so I think uh we can start with the session guys am I audible to all of you please do let me know in the chat if you can hear me then I'm waiting for your replies tell me guys uh if I'm Audible and visible to all of you please do let me know so that I can start with the session yes okay great so very good afternoon to all of you so today uh we'll try to discuss uh geni related interview questions so yesterday actually I completed the machine learning and I started with a deep learning so in today's session actually I'm thinking about about the Deep learning so couple of question related to the Deep learning like computer vision and the core NLP and then I will move to the generative AI so I hope guys you have seen the uh thumbnail and the title of the video so yes going to discuss the question of the generative AI we'll discuss kind of road map as well so what will thing actually you should learn whenever you are going to start with the generative AI okay and apart from that what type of inter questions you might get whenever like you are going for the Gen interview that what type of interview questions you might get definitely we'll try to discuss that as well so tell me guys uh if everything is fine from my side then please give me quick thumbs up so that I can start with the session tell me guys fast waiting for your reply in the chat so please do let me know yes okay great then uh fine so first of all let me give you the quick uh overview of the dashboard itself so everything uh we are updating over the dashboard if you want to get every material right uh whatever material I'm discussing here inside the live class inside the live session so you can download it from you can download it from the dashboard itself so so far we have uploaded the question right so related to the python related to the statistic and the Machine learning so inside the resource section you can go and check so let's say here we are talking about the previous class guys deep learning interview question so this particular question so first of all let me show you that how you can navigate to this dashboard so if you want to navigate to this dashboard this uh data science interview Professional Series dashboard so first first of all you will have to go through with your description description of the current video the video which you are watching on YouTube right so the video basically which you are watching on YouTube you have to go through with the description of that so once you will go through with the description here you will get a link for the dashboard right so just click on this link then directly you will be redirecting to the dashboard itself here is a dashboard now once you will open the dashboard guys first of all you should have access of this dashboard so it is completely free no need to pay anything for this interview series and for the guidance regarding the interviews and all it's completely free here you can see right so then you can enroll you can click on this enroll now and uh finally you will get your dashboard so this is a dashboard now all the recording assignment quizes everything you will find out over here and you can download the resources as well from this resource section so whatever PDF whatever resources I'm discussing throughout my session every type of resources we are updating here inside this resource section so please go and check with the resource section and try to download it now you will ask to me sunny uh the resources is not available from the previous class so guys this a PDF this one right so once you will click on this machine learning interview questions in machine learning interview questions actually the same PDF uh the same like a question set I have used for the Deep learning as well because inside that itself we have a deep planning question right so once you will click on it so here you will get this uh drive now you can see the interview question PDF 1 2 3 and here you can see the data science interview question so 30 days data science interview question so this is the very famous interview question set from the Inon uh if you will go and check with the Inon GitHub or so definitely will get this 30 days interview question of the data science preparation data science interview preparation you can get that also from here itself from the resource section so just uh enroll to yourself and try to take a leverage of these resources so tell me guys uh regarding dashboard resources and quizzes assignment everything is perfect yes can we start with the session now so just give me quick thumbs up so that uh we can start with our main agenda tell me so I'm waiting for a replies in the chat I think 11 people are watching to me currently so yes if you're watching to me then please do let me know in the chat okay so momas is say yes sir you can start now uh what about others okay great so fine uh let's start with the session so in the previous class I started with the Deep learning interview question right so we discussed couple of interview question related to the Deep learning now first of all let me give you the quick overview of the artificial neural network then I will come to the convolution neural network and I will show you the interview question related to the computer vision that what kind of interview question you might get related to the computer vision uh if you are going in an interview that what will be the question related to the projects everything we'll try to discuss over here and don't miss out tomorrow's session where I will discuss about the resum building and all so I will give you some key points key idea that what all thing you need to mention whenever you are preparing your your uh whenever you are preparing your resume so definitely that uh tomorrow session is going to be a really interesting one today I will try to complete this uh deep learning as well as the generative AI if not then definitely we'll take couple of more question in tomorrow session and then we will start with our resumer building because during this data science uh like uh during the interview preparation actually you need to keep in your mind that interview Co this resume also plays a very important role until your resume is not going to be select then definitely uh no one will take your interview and uh this preparation and all everything is based so yes uh this resume actually uh it's a very important part of the interview preparation series and definitely Ely I will try to discuss in tomorrow's session the key points are all related to the resume so fine uh let's start let's begin with the session so first of all guys let's try to discuss uh couple of thing related to the neural network and then I will come to the main interview questions I have uh many more questions and I will show you all the possibility basically because during my experience or during my teaching experience mentoring and all I uh went through with the different different resources even whenever I given the interview so all those uh like interview question basically I capped at somewhere so definitely I will try to share that also with all of you now first what I can do so here let me share my board and there itself I can give you the quick overview related to this deep learning topic so that you will get some idea that what all topic you need to learn whenever you are going through the deep learning and what type of interview question people might ask you so yesterday itself I told you that uh deep learning actually we can divide into the various section uh so for the first and very common section which is called artificial neural network so definitely they will ask you some fundamental question from the artificial neural network then the second is what convolution neural network in this one you are going to main you are going to deal with the images so it's it is having one more Branch right the branch itself is called CV computer vision right so here actually we generally uh talk about the images we try to like solve the different different challenges related to the images then the third is called NLP so here is what uh sorry RNN and from here itself you'll find out one more Branch okay that got that is called NLP itself and here you'll find out the generative AI right so here itself in this particular section you'll find out this generative AI generative AI so this is a extension of the NLP itself the llms which you can see right currently the llms you are seeing large language models right so it's a extension of the uh like the different different model which was already there inside the NLP getting my point so mainly here was a Transformer and it was the extension of that now a couple of more topic like reinforcement learning and all I think that is also part of the deep learning so reinforcement learning is also there gains is also there so gains actually it's a part of the deep learning itself gain generative adversor Network for data generation got it now guys uh the topic wise I was discussing different different topics so the first was the introduction of the neural network so here I can write the couple of topics okay this is the five main topic inside the Deep learning itself now let me write couple of topics over here which you need to focus whenever you are preparing for the uh deep learning interview so the first one basically introduction of the deep learning introduction of the neural network don't worry I will give you the questions as well first let me write down the topic over here so introduction intro of neural network intro of neural network okay so actually what is neural network what is artificial neural network right so these type of question definitely you might get the second is what introduction of neural network introduction of the deep learning right and the second is what the second is called uh perceptron basically so Perron is a very important one so perceptron okay perceptron now here you can see the next one third one basically so the third topic is going to be activation function right activation of function then uh we have the loss function so here let me write the loss function okay so loss function loss function right now the next one basically uh the third one fifth one is going to be a back propagation back propagation right back propagation back propagation now the next one topic next topic here you will find out the uh basically method mathematical equations right all the mathematics behind this particular topic derivations and all everything then uh basically the next one which you will find out uh that is called call backs call backs is very much important so why where we need to use the call callbacks and all okay so what is a call back then early stopping right early stopping so early stopping is one of the topic so which you need to focus and definitely it's very important whenever you want to like enhance the performance of the neural network right then uh the next one basically normalization regularization so here the next topic is what regularization regularization then the normalization normalization is one of them so here is what next is normalization okay normalization then weight initialization weight initialization weights initialization weight initialization and then you'll find out the next one uh that is what so weight initialization hyper parameter tuning hyper parameter parameter tuning so hyper parameter tuning is one of the important one so what should be the learning rate okay how much EPO you should like uh how much epox you should run and which like how much ax you should run how much hidden layer you need to take care of right so basically for the better optimization and all so these are the couple of topic which you should always learn whenever you are going for the artificial neural network and definitely people will ask you the question from here itself now our next topic basically back propagation so here this optimizers optimizers different different type of optimizers right optimizers okay so different different type of optimizers and all so introduction percepton activation loss function optimizers call backs early stopping regularization normalization weight initialization hyper parameter tuning and all and definitely you can crack any sort of an interview related to the Deep blooding right if you have a grip on this particular topic right fundamental grip on these particular topics right now the next one is what so here I think this part is getting clear to all of you now the next one is what so here we need to understand the fun fundamental differences between the ML and DL So Yesterday itself I told you whenever we are going to talk about the differences between the ML and DL So based on the various point we can do that because from here directly I will I will show you the interview question and I will jump to the next topic that is going to be a computer vision and then finally generate and NLP okay so uh here the next one is what so if we are talking about the we are talking about the differences between what between the like ML and DL so on the five basis actually you can do that so here let me write down the differences ml versus DL so ml versus DL okay now the differences wise the first difference basically you can do based on the data the first difference you can make Based on data so here actually the data wise we are using uh hug data set okay generally we pass the used data set and generally we deal with the unstructured data also like images videos audios right now the second is what send second is the training time so here training time is also there so let me write the next one that is going to be training time so training time is there okay data is there training time is there right now next one is what Hardware dependency so definitely uh Hardware dependency will be there so Hardware dependency Hardware dependency okay hardware and here next is what Hardware dependency see now the next is what tell me guys the next is uh the next could be here tell me what you think so here the next uh could be the feature extraction and feature selection right so feature extraction and feature selection right now here the next uh is what so interpretability of the model so interpretability of the model or explainability right so here interpretability is a little uh like complicated compared to the machine learning model so based on that also you can make a differences inter inter tability getting my point so here this is the next point so based on these particular points definitely you can make a differences between this machine learning and deep learning so yes definitely I can summarize this thing now here if we are talking about the data right so the data basically here uh like uh here we required uh like if if we are talking about the classical machine learning algorithm so there we are passing basically tabular data which is containing the either a numerical or categorical feature okay but here in the Deep learning so specifically we are dealing with the unstructured data because we have those technique those algorithm which which can deal with the unstructured data right the second thing is what training time definitely this uh machine learning deep learning model actually takes huge time because uh here the number of meter is very high compared to the machine learning model so definitely it takes huge time whenever we are going to train that now the third thing is what Hardware dependency this machine learning model is not having any hardware dependency means we can simply train it on a CPU as well means on a lower configuration but if we are talking about the Deep learning model so because of the complexity of the model so definitely it's going to take a lots of time so the processing uh processor and the ram right the configuration of the model should be very very good right so so that's why like researcher has introduced the gpus the cap because before the before 2012 okay before this imet competition people were not using the gpus in the in the like field of the deep learning but uh like this Jeffrey Hinton actually he he came with the he came with one like CNN architecture with the I think the name was a v maybe so I if I'm not wrong V or Alex net so he has introduced that particular architecture along with the G us and the processing time was very very fast in 2012 itself so in the image n competition he has participated and even on the competition as well along with a good accuracy so definitely Hardware dependency is there with respect to the Deep learning model but in terms of machine learning model it is not there now the next point is what features engineering or feature extraction of feature selection so whenever we are talking about the like uh whenever we are talking about the like deep learning model right so here actually automatically they can extract the feature from the data itself so uh let me give you one example so here very good example uh related to the resume you know about the resume resume right so resume now here if I want to classify the resume resume of the person so what I will have to do tell me so here if I want to classify the resume based on the ml based algorithm so what I will do tell me I will create a manual feature what I will do I will create a manual feature so F1 F2 f 3 fub1 FS2 right fs3 fs3 then F4 and based on that basically we are going to classify that whether this resume belong to the like HR department or operation Operation department or sales marketing Mentor developer right so based on that based on this particular feature we can classify that so manually actually we are going to create a features over here but if we are talking about the de right so here it can automatically extract the feature from the uh resume right so I I think you know about the computer vision like CNN architecture so it automatically uh it automatically like uh extract the feature right automatically the filter and all is going to be learned right and definitely we can extract the like data and feature from the resume itself and based on that we can classify so it belong to which particular class so this automatically feature automatically feature like selection is there feature extraction is there with respect to this deep learning now here guys if we are talking about the interpretability so it is little hard compared to the it is little hard compared to the like machine learning model so the complexity of the model is very very high because of the like neural network and we are going to make a multiple connection between the input and hidle layer and if we are going to take a multiple hidden layer so the interpretability of the model is going to be really hard so definitely on this five terms on this five fundamental terms you can decide that right now the next question basically which I told you in my previous class so here how you can increase the power of your neural network if you don't know about the neural network so you can go and check with the previous session there I already discussed about the neural network that what is a neural network and uh what is a forward propagation work propagation right parameters loss Optimizer everything so if you will uh check with that session just check with the last part I think uh between uh like before completing the session so within 15 minutes actually I try to explain that so you can check with this particular part if you want to understand the uh like neural network if you want to understand about the neural network now someone is saying not hint is was it was a Yan look no I think uh the person was the Jeffrey inter if I'm not wrong you can check with the Google so here if you're going to search with a image net so let me show you that in front of you only I can search just wait uh here what I'm doing so here I'm searching image net competition image net competition uh I hope you will get the entire detail over here so see it is saying that uh about the image net competition so first of all check the full form of it image net large scale visual recognization challenge now Workshop Download update stion okay it's a detail of the image net from the official website so imaginate competition 2012 2012 Yan L actually he introduced the first model of the CNN that was the Linette in 1998 maybe if I'm not wrong you can check so that was a CNN architecture and they have trained on the amist data set they took the amist data side and they have trained the model on the amist data set and that was the first model lenet and the M I think the name actually you can match Yan Lin right so the Lee is coming right I think based on that they have uh they have named the architecture as uh linet basically so now you can check with the image net competition of 2012 so who was a winner and all L using the data set okay Workshop schedule introduction data task one classification classification tation no here is not there on contion network okay Alex net was the architecture see I was saying here Alex net was the architecture the imag project database use object research 14 million images indicate the 1 million imag more than 20 20,000 category URL since software program significance of De history of the database okay data set was there subnet history of the image challenge tot 5 and here 2003 I a top five a deep learning conv next couple of year so 2014 winner so winner winner winner winner winner Alex net Alex knet was the winner guys and I think it was developed by the Jeffrey Hinton if I am not wrong you can do the more research on top of it okay let's uh discuss later on about it otherwise uh we are going to leave the topics and all okay now here alexnet was the architecture which was introduced inside the in this particular competition imaginate competition now I hope this thing is clear to all of you data training time Hardware feature extraction interpretability and all now the next thing is what so here the next thing is how we can improve the performance of the neural network right so can you give uh also please give some logical question being asked yes yes definitely we are going to discuss that I will show you first of all let me introduce the topic so that uh no one will feel that this question from where it is coming because everyone from the different different like places now so some of you don't know about the Deep learning some of you don't know about the artificial neural network and CNN so that's fine now here if we are talking about the uh like the interview question so how we can improve the performance performance of the neural network so neural network guys if we have talking about the neural network so that is a very important question and in most of the interview basically people might ask you this type of interview question so the first one basically you should uh focus on the hyper parameter tuning right hyper parameter tuning of the model hyper parameter tuning and that is very much important now here in the hyper parameter tuning there are many things basically which comes now whenever you will look into the libraries like image uh this tensor flow kasas or maybe P toor there you will get like a lots of hyper parameter so no need to look any lowlevel Library like tensorflow or like py toor you can look into the Kass basically there you will find out the like a inbuilt classes you just need to call that right if you don't know about the Kass so Kass actually it's a uh UI on top of the tensor flow itself so there you don't need to write any lowlevel functionality you can directly call the different different method and directly you can create a different different object along with different different parameter basically okay so hyper parameter tuning is very much necessary so in that you will find out the number of notes right how many notes actually you need to select you'll find out the number of uh layers right how many layers basically you need to select layers notes and uh basically right so how many box so these are the very important thing epox will be there now uh here hyper parameter wise basically so uh the next one Optimizer so Optimizer also plays a very important role and you should always go with the advanced optimization technique right always uh you should go with the uh like Advanced optimization technique like uh Adam Optimizer and the different different variant of the Adam Optimizer nowadays actually you will find out so if you will check with our different different website like uh uh papers with code just go and check with that and you will find out lots of like stateof art optimization technique and definitely you can uh go with that and you can like uh build your neural network but nowadays actually everyone is using the pre-train model or they are trying to do the tuning of that but yes this fundamental it should be clear because uh in interview they ask the question from there itself but in real time you are just going to be import the library from somewhere and you are working uh on the project right so that's the reality but on uh like uh for reaching to that level the fundamental should be clear the next one is what the next one is the next thing Optimizer is very plays a very important role okay now regularization technique reg regularization technique also very much important regularization and you should always focus on that regularization technique now after the regularization there is like a normalization technique so for the fast processing and for the better performance of the neural network model you can perform the normalization also and that is also very very important normalization is there regularization is there getting my point guys so you should always focus on top of this now regation wise L1 regularization L2 regularization L1 plus L2 regularization right solar storage and all these are the regulation technique now here the next one is what so here the next one if you want to get a better performance of your neural network so couple of more point I can discuss over here the next one is going to be a weight initialization technique so definitely if you want to improve the performance of the neural network so you should always focus on the weight initialization technique which weight initialization technique you are going to select right and that plays a very important role whenever you are doing uh whenever you are going to train your model now the next one is what so here it could be like uh uh early stopping early stopping also so unnecessary your computation is not uh like unnecessary you are not burning your resources so yes you should focus on the r stopping and all then uh you should focus on the best size right best size how much best size you need to be passed whenever you are going to train your model so there are a couple of ways by using that you can increase the performance of your neural network and this was the very common question so that's why I discuss it over here now EPO note layers like B size early stopping weight initialization it plays a very important role whenever you are dealing with the neural network right so selecting layer and nodes is a trail R error am I correct Trail R error selecting layers and noes no it's not like that I didn't get a question if you can be more clear then maybe I will get it for l net it's a Jeffrey intern yes correct so for alexnet it was a Jeffrey intern and I was right now I think this thing is clear to all of you coming to the interview question let me show you the interview question based on that and then I will come to the CV computer visign based on interview question so first of all guys uh here what I'm doing uh let me show you the interview question based on this one so here I kepted already in my GitHub now I am opening my GitHub basically um just a second okay this one guys so here you can see this is what this is my interview question so part two part three I have three sets basically so in the set two automatically you will get all the interview question whatever I have discussed so far now apart from this one uh like uh if you will go and check in the resource section basically you will get the interview 30 days interview question also of the uh from the Inon right so like actually this uh 30 today's interview question we are including everything over here inside this particular question right deep learning machine learning statistic time series everything every like a basic fundamental thing you will find out over here itself so this interview question along with the answers but here now let's let's focus on the Deep learning itself so from here the Deep learning question is going to be start I hope it is visible to all of you this one this particular question so uh let's discuss now so first of all guys here you can see how will you explain uh the training of the neural network can you visualize the neural network what are the different weight initialization technique right why do we neural network instead of the state forward mathematical equation means machine learning equation okay now here why do what do you understand by back propagation in neural network can I solve the like tabular data by using the neural network right classification problem statement so give the L uh give the C list of the cost functions now here can you correlate the biological neural network with the neural net biological Network with the artificial neural network so here see the interview question now give me the list of 10 activation function so these are the interview question from the basic and fundamental now coming over here now explain disadvantage of the reu function right what might be the disadvantage of the reu function now how do you select number of layers number of neuron in a neural network so it's a hyperparameter and it's a based on an experiment only now have you ever designed any neural network architecture by yourself so yes definitely uh like you can answer this type of question can you explain the swis function now here what is the difference between batch mini batch and the sastic gradient explain five best Optimizer right now here you can see all the like till this uh 116 actually you'll find out all the interview question related to the fundamental artificial neural network here you can see I hope it is visible to all of you over the screen itself now uh I think this is fine and here it is not a big deal basically and don't worry question and answer everything is there so this is just a question in this set particular sets and definitely you should work on it if you want answer basically so definitely you can go and check with this uh 30 days interview question preparation here you will get the answers as well along with the questions right so here you will find out some other questions and this is also I'm going to be upload in my GitHub and it will be a it is already available inside your resource section so from there also you can download it now here also you find out the fundamental question related to the Deep learning right so this I'm not introducing as of now so here I'm like stick with this particular question but yeah just for the reference right just for the like uh variety of the questions and the answers I'm going to share that also with all of you but here you can see this is the precise question based on the topic which I was discussing over here right I will give you the additional material as well which I shown you right now currently right so yes definitely you can go through with that material also and you will get a variety variety of the question and but most of the question 99% of the question I'm trying to cover it over here itself now the next one is what so here the next one is related to the convolution neural network so I will come to the convolution neural network and before that uh I will show you something else so here basically I have the interview question on top of the computer vision as well so 100 computer vision interview question is specifically for the computer vision if someone wants to prepare for the computer vision right specifically for the computer vision and all definitely you should read this particular interview question now everywhere actually you will find out the like large model so Transformer based model and yes it is applicable this Transformer based model actually those model is applicable for the computer vision use cases as well like detection segmentation and all getting my point so yes uh that will be my main agenda after discussing this fundamental right so this is the fundamental thing uh Vision related to The Vision CNN and the RNN and then generative models right from where basically you can read more about the generative model generative Ai and uh the Transformer based model on top of the Transformer also there are many more things is coming now so yes the those are beating the Benchmark of the Transformer Benchmark means what Benchmark means on top of the like data right the The Benchmark data basically what is the per performance of that particular model in a different different scenario in the different different situation right so those Benchmark is bitten by like many of the model I think you have heard about the Mamba right so on top of the Transformer it is working well and yes U I think uh most of the like technology nowadays is coming so but yeah transform based model it is it is outperforming among all those model and definitely like uh like I will give you some sort of a guidance related to that so so here first of all uh I think related to the Ann everything is s out okay so I given you the enough amount of understanding here again I'm saying I'm not going to discuss in detail mathematics and all so for that maybe other time some other time but yes you should take a reference of this interview questions if you want to crack the interview a fundamental of it and definitely 90 to 95% question like people ask you from here itself got it now what we can do here in today's interview question there are question only till day 19 the resource section till 90 till 90 okay I will check Min if it is not there with the itself so it will be available can hyper parameter tuning for deep learning model selected by any automated method rather than manually trying different yes so manual basically you are using the libraries now you can use the cross tuner right or any other libraries like optuna hyper op and all by using that you can select that right just use a classer automatically each and every experiment it will perform in the back end and it will provide you the best set of the hyper parameter you don't need to select manually right so if like you have a combination where uh you have around if you have hyper parameter and the combination is around let's say 1,000 so 1,000 time you will train model manually no it is not possible you have automated library you just need to pass the value and automatically it will do everything for you getting my point I hope it is clear to improve the Transformer how many notes and layers is up to us we can improve by adding more am I correct is there a specific strategy no how I can tell you the specific number if I will tell you use 10 notes and 10 H layer so is it true for every scenario tell me is it not true it's the experimental thing right it's experimental thing and every uh instead of art model basically whatever state of art model you will find out so in that you will see that uh like clearly they have they are mentioning in the research paper also so if you will look into the Transformer model they are they they are taking basically self attention right so multi attention they are they are taking eight multi attention and on top of that four six like six stack of the encod and six St of the decoder there also we have a neural network now so six neural network and the six neural network decoder side why they are using six why not other so it's a experiment based thing and definitely there is no hardcoded Val for that fine so let's try to discuss about the CNN so first of all let me give you the idea uh behind the CNN and then I will come to the interview question of that great so first of all let me discuss about the CNN architecture so we are talking about the CNN guys so what is the full form of the CNN so the full form of the CNN is convolution neural network right convolution neural network so in the convolution neural Network what we are doing so here is what here is my image so this image actually I want to pass to my neural network so first of all you should have understanding of this image okay so this image actually is nothing it's a collection of pixels so collection of pixel right now this pixel actually is having some sort of a range let's say the range is from 0 to 2555 in black and white image actually you will find out only two channels two channels now in colorful image actually you will find out the you will find out the three channel right so here colorful image colorful image so there you'll find out the three channel right so three Channel now here in the black and white actually uh in the black and white you'll find out only two type two pixels not two channel so there will be only single Channel let me write it down here again so here in the okay fine so we are talking about the black and white here only you will find out two pixel okay binary pixel either black or white so black means what black means zero and white is representing to the 255 now here if we are talking about the colorful so here there you will find out r g and B three channels basically and the range between from 0 to 255 now here in the black and white either this one or this one right so this is the basic anatomy of the image now what I have to do I have to pass it to the neural network so here is what here let's say is my neural network I have to pass it to my neural network but I never pass directly this particular image to my neural network right neural network so here is what here is my neural network let's say this is what this is my fully connected neural network I can pass it directly also I can pass this image directly also to my NE network but we always pre-process it right and that pre-process is called convolution convolution so here actually we have neural network we have neural network along with what along with the convolution right convolution plus neural network in the convolution actually there are certain steps so the first basically what we do we always multiply it with the kernels right kernels or this is also called feature so this kernel actually it can be any sort of a metrix let's say 3 cross three Matrix 5 cross five Matrix and this steps is called convolution okay we are going to multiply it this step itself is called convolution okay convolution and it is also called the feature extraction feature extraction we are extracting the feature from the images so like this this this one this like matrixes which is there right this three cross three matrices and all so it's a learnable parameter it's a learnable parameter right automatically is getting learned learnable parameter so we are assigning the weights okay we are assigning some sort of a value and in the back propagation actually it's trying to learn themselves this particular parameter then what we do after the convolution actually we perform the railo on top of that we perform the railo function on top of that on top of the convoluted image and then after the ra we perform the pooling operation now pooling actually we do it for getting the specific feature from the images right so there can be a different different type of pooling like minimum pooling maximum pulling right here can be like average pooling different different type of pooling actually you will find out so minimum pulling maximum pooling average pulling and then finally we are doing a flattening so flattening is nothing basically it is representing to the array right 1D array so this finally after the pre-processing this after the pre-processing we are going to convert this image into the 1D array this particular image into the 1D array and we are passing to the neural network so this is what this is my input layer right here is what here is my input layer this is representing to my hidden layer this is representing to my hidden layer this is representing to my another hidden layer hidden Layer Two and this is what this is finally my output layer so I hope you're getting my point so this is complete process of the convolution and this architecture has been introduced by the yan lukin yan luin right lukin and if I'm not wrong so this is the spelling uh in 1998 in 1998 he has developed an architecture the architecture name was the lanate okay lenet and he has developed this particular architecture with respect to this amist data amist data okay handwritten images Amin data so I hope this thing is clear of the convolution neural network now guys here this a parameter right this a pooling right uh So based this number of parameter based on this pooling and all we can design a several type of architecture we can design the several type of architectures here what we can do tell me we can design the several type of architectures so here guys so if we are doing a changes right where inside the inside the number of Fe inside the number of kernels right the number of of feature matrixes or if we are changing the pooling right if we are going to use a multiple convolution if we are going to use a multiple pooling right so we can make a changes inside those parameter and based on that actually we are we have generated a different different architecture we have generated a different different architecture so here I can show you those particular architecture where uh let me show you over the Google itself so what I'm doing here let me search about the okay Kass CNN so here I'm searching about the Kass CNN architecture now once you will search it so guys here you will get maybe all the CNN architecture right so let me show you that if I'm not wrong so definitely I will get that okay so just a second grass CNN architecture um image net simple amist CNN customization overview or text I think they will give you the list of the architecture grass all CNN architecture yeah guys so here you can see this is all the architecture based on what based on the changes in the kernels okay based on the changes in the number of features feature is what nothing kernel so how many time you are doing a convolution and all so just just look into the number of architect number of architecture right we have like lots of architecture lots of architecture like exception is there VCC is there VCC 19 is there rest net is there rest net or different different variant of the res net right Inception is there mobile night is there dens net is there Nest net is there okay efficient de efficient net right so many more architecture so let's say if you want to check the reset so here what I will do I will check with with the rest net architecture so rest net architecture so you will find out the uh like uh the different different like a the number different different number of convolution and all right so just just look into this particular architecture that how many convolution layer is there how many convolution layer is there inside this architecture right how many time they are doing a convolution regarding the different different parameter so if we are changing the number of parameter So based on that basically we have generated a different different type of architecture and for regarding every architecture you will find out the research paper getting my point now you can design your own custom architecture as well now here guys see this is the raw CNR architecture which is doing the image classification what it is doing tell me it is performing a image classification now we have one more task on top of the image now right we have one more task so here the task name is what the task name is uh object localization so actually there we are just trying to find out the location of the object inside the image right now here uh okay I will come to that first of all try to understand the CNN architecture so you can go through with any sort of architecture let's say here I'm searching about the efficient net so in the efficient net so efficient uh net architecture so see the number of images and all okay so there there will be many many more variants right many more variants inside the architecture so how many number of Channel okay width of the channel right and here you can see the depth of the channel uh depth of the channel means what features feature if we are talking about the channel means feature right now here you can see the wider deeper wider and all right based on that based on the changes and all they are going to create a different different architecture now let me show you one more architecture which has been introduced by the Google itself the architecture name was the Inception net so here Inception architecture now inside that you'll find out something different so they have introduced this parallelization okay parall basically they were trying to extract the feature in the ret actually they have introduced the uh they have introduced the like uh this one they have introduced this residual connection okay and in efficient architecture in the efficient net so like parall basically on a different different level they are trying to extract the feature that's it at the end everything is about the feature extraction because the image image I told you I given you the anatomy of the image so this image can be a different different one in a different different scenario in a different different domains right so the image can could be anything right so the image uh there can be M image there can be like any uh like software image there can be any like motherboard image okay the image can be a like space image right or in a different different condition also so pixel varies actually in each and every situation so there can be a standard definition high definition right or high quality maybe uh like infra image or like different different type of of image right so this image actually it varies based on the like based on the use case based on a problem statement so definitely the number of pixel right or the number of uh like comp the kind of complexity basically you might get a different different one so for extracting those feature in efficient manner we are trying to design a different different features getting my point you can do your Mo research on top of this images right and you'll find out lots of images now let's see I'm uh like standing like this and you are watching to me right now someone is taking my drone shot means someone is taking from the like let's say from the 100 ft or maybe 200 ft so in that case the image will be different now here I'm standing in front of you in this particular lights now let's say if I'm going to switch off the light so in that case uh the image uh basically it will be a different right so image condition could be a different in a different different scenario and based on that we have to extract the feature uh from the images itself and that's why they have designed the different different architecture getting my point and the main and the native application of this CNL architecture the main and the native uh application of the CNN architecture was nothing it was a image classification they are trying to classify the images right now the second thing is what so here let me give you some sort of idea related to the other task as well and then I will come to the final interview question I will show you that so here I'm just trying to clarify the basics and then interview questions now the next here uh let's say this is what this is my image so here what I'm doing let's say inside this image I have one object and this object is what let's say there is one person so what I'm trying to do here so here actually I'm trying to classify it so whether this person is sunny or maybe someone else right so this is simply called classification problem classification problem classification problem okay now here let's say here is a image now the same image right whether I'm trying to identify right this person is sunny or maybe someone else okay maybe or someone else so here I have two classes so sunny or someone right so it's a simply classification problem now let's say this is what this is my person now this is what this is my person now here I want to localize this person okay I want to check the uh what I want to check guys tell me I want to check the position of the person position of the person so this is called uh object localization object loc lization now here what I'm doing so here what I'm doing so I'm going to trying to find the position of the local position of the person now here what I have I have the image right now inside this particular image let's say we have two object so here what I have I have two object so this is my one object and let's say this is my another object so what I can do I can draw one kite so something like this right so here first I will do I will do the object localization and on top of that I will perform the classification so this is called so this is my uh position of the image this one sorry position of the object so this is called localization localize ation okay Lo this is what this is a localization and here I'm localizing it so I'm just trying to find out the position of the kite localization localization and then on top of that I'm performing classification so this is simply called the object detection means I'm trying to detect the object inside the image object detection so localization plus classification so localization plus classification getting my point here so localization plus classification and for this also the same thing and that is nothing that is called the that is called that is called the object detection so here also in the object detection actually we have a different different framework as you know right so here I can write the different different name let's say here is what uh maybe here is uh you know about the faster rcnn okay CNN family so faster rcnn is one of the very famous model now YOLO is one of the very famous SSD is there right and t4d detron couple of other framework and efficient date and all right so these are the framework now this this this architecture is nothing actually this architecture okay this is also a CNN based architecture only right so if you will look into this faster CN so it is nothing it's a CNN based architecture because here also we are going to perform this classification right now here one more thing is there localization so here along with the classification actually you'll find out the localization localization actually it's nothing they're trying to identify the position of the object inside the image right so for that actually they are uh taking a different different approaches they are just taking the region interest region of interest from the image okay and on top of first they are selecting randomly and then they are regressing on top of it means by learning only learning the coordinate mean means first they are trying to uh simply what they are doing see uh here Roi actually you will find out Roi reason of Interest so let's say this is my image this is my image so randomly I'm selecting a region from this image let's say this particular region now uh here this is the coordinate of my region and here is what here is my I'm assuming that my object is here so let's say this is my object my object is present over here and this is the region basically which I selected region of Interest now what I'm doing I'm regressing on top of the coordinate okay I'm ressing on top of the coordinate and I'm trying to find out the best possible coordinate where my object is present this one okay so along with the classification this are different different concept is involved okay Roi and all and on top of that we are regressing but at the end basically we are trying to find out the correct reason correct like a reason or the correct object means we are trying to localize the object we are trying to localize the object and then basically we are trying to classify that classification will tell me so classification will tell me okay so who inside who is inside the object so let's say there is a sunny okay so one person let's say in the next one uh there is one cat okay let's say one uh other object let's say this marker or maybe laptop or any other person so whenever we have multiple object in single image so we always per perform the object detection to identify the to identify the tell me to identify the object from the given image right so whenever we have a multiple object inside the given image we always perform the object detection getting my point now here also we have two variants of it two variants of it see this is also supervised learning this is also supervised learning means here also you will find out the so first like uh we are doing a calculations right randomly we are selecting something so calculation and this calculation basically we are trying to identify whether it is correct or Not So based on what based on the loss function and then finally we are trying to perform the uh like optimization okay based on the uh like a different different optimization technique so this is also a classif this is also supervised learning this is also a supervised learning but here actually we have one more Vari so instead of detecting object what we are doing you know so we are going pixel to pixel here pixel to pixel and we are creating one boundary surrounding to the object and that is called the segmentation right and the process will be same you can look into the uh you can look into any sort of architecture I can show you that just wait so here what I'm doing so here uh I'm showing you the YOLO architecture and definitely by seeing that you will get to know about it so what I'm doing so here actually I'm going to search the YOLO architecture so YOLO architecture now you will find out uh in the YOLO architecture is nothing it's completely convolution only it's completely convolution any any sort of architecture whatever architecture basically you can see okay this is the YOLO architecture I think it is visible to all of you if not then let me Zoom uh let me zoom in and see this is the YOLO architecture so it is what it's a convolution only but here certain thing is involved for what for the localization I want to make my local localization as much as better so I will use the different different technique I will use the uh like Roi technique reason of Interest I will select that and then for the evaluation actually there is a uh like uh intersect over the Union right so that is a technique just on top along with the classification only we are trying to classify the object only inside the the image itself and that is what there is a object detection that's it now noways this thing we can perform by using the Transformer based architecture also so here let me show you see first of all let me show you some other architecture which is very very famous so here I'm showing you the faster rcnn I think you have heard about this faster rcnn so faster rcnn architecture and you'll find out some latest arure also so see one of the architecture which I recently seen that is what efficient debt so efficient debt by using this architecture also you can per from the object detection this is the architecture and here you will see it is nothing it's a uh like convolution only see everything convolution is happening nothing else convolution so convolution actually is a fundamental unit convolution is a fundamental unit of the computer vision convolution neural network is what it's a fundamental unit of the computer vision now here guys see uh convolution is fine efficient dat is fine faster RCN and architecture let me show you the faster RCN architecture see this is a foster RCN architecture what is happening so here's a reason proposal Network which is proposing the reason mean which which one should you reason for identify the object and on top of that you can see so here is what uh convolution neural network that's it so here is what uh this is my feature map means uh it is trying to exct the regon from that this there is one network who is proposing the reason and on top of that and at the end we are going to perform the classification means convolution neural network that's it and that is nothing that is faster RCN I hope you are getting my point now you can look into the other architecture also like SSD so this is also very famous architecture see what is happening over here inside that only the uh SSD architecture let me show you the SSD also I think this is the SD architecture see they are using the convolution only in backend nothing else so I hope this thing is clear to all of you now we can look into the interview question that what all type of interview question you might get if you're are specifically going for the computer vision interview now if I'm searching about the Transformer Vision okay Vision Transformer Vision Transformer right so this uh particular architecture actually it has been built on top of the Transformer for the object detection use cases right object detection object segmentation and maybe image classification also you can perform by using that so here how it is happening see so just go and check with the paper and here actually you will find out every detail inside the paper itself along with the architecture and all everything uh they have provided you over here itself so this paper actually they have introduced in 2021 and here you can see an image is a what 660 cross4 Transformer from image recognization at scale now just read this particular research paper and here they have given you the architecture as well I hope it is visible to all of you so how it is working see first linear projection and then uh like embedding and all something and then finally they are passing to the Transformer encoder and this is a like high level overview of the trans this particular Vision Transformer okay now you can uh check with a multiple research paper multiple model and yes Transformer based model also you can use for the vision based task is it getting clear to all of you yes or no tell me guys fast is it clear if is it then please give me a quick thumbs up so that I can uh I can proceed further yes correct Transformer architecture you are right anyone if you have anyone if you having any sort of a doubt you can ask me in the in the chat and then I will proceed with the further interview questions okay okay so I hope uh everything is clear so far now let's look into the interview question so first of all let me show you some fundamental interview question on top of this uh CNN and then I will come to the computer vision okay so what I can do so let me show you the fundamental interview question interview question set two and here you will find out from here basically from question number 117 the fundamental of CNN is going to be start explain the technique for doing data argumentation in deep learning so can anyone explain me the meaning of the data argumentation what do you think guys what is the meaning of the data argumentation do you understand what is the meaning of the data argumentation yes or no tell me have you heard about this data argumentation ever question to all of you who all are watching to me then please do let me know guys fast creating different version of the data correct getting more information from the data okay getting more information from the data no this definition is seems correct creating different versions of the data means we are argumenting the data we are enhancing the information right so we are trying to generalize the data generalize means so from everywhere it can capture okay so uh from every like angle uh it can recognize the image okay we are argumenting it we are enhancing it right I think this argumentation term is very much like common right retrial argument system have you heard about the retrial argument system is based on that only means based on the information from the database uh which we are collecting from the databases we are passing to the LF and it is enhancing the Generation generation of the data so in a similar fashion if you want to make a more capable model so that it can capture all sort of a pattern or the larger information in that case we perform the argumentation okay so argumentation means let's say if I'm uh like looking into the camera and here standing like this so definitely you can uh recognize me this is s but now let's say if I'm standing like this so maybe you won't be able to recognize to me if I'm watching to this side in this direction now let's say if I'm extending like this okay now I'm looking into the W so maybe you won't be able to identify me because of what because I'm looking it at this board so in a similar fashion if I want to make my model too much capable so that it can detect me from every angle like this okay like this there or there so for that we try to give the different different variety of the image like uh where we are going to tilt the image where we are going to like uh performing the horizontal scaling vertical scaling or different different gradient actually we are trying to apply okay or we are removing the corners or we are trying to move the objects and all right so different different type of thing we are going to perform with respect to the data and that is called the argumentation argumentation I hope this is clear to all of you and and uh here like this thing is clear to all of you now so in llm do interview ask the right quote from scratch yes uh it depends on the interviewer basically so let's say if you're are going for the interview and uh that day interviewer is not in a good mood so definitely he can tell you okay so please write the code for the Transformer I want the code in a scratch from scratch just write to write the self attention by using the python so it depends on the day and all but you should be prepared with each each and every expect right so at least try to code whatever thing you are learning because everything at the end you are going to implement by using the code only so let's say if you're learning any sort of a concept any any any type of concept so definitely you should code it later on at least the common one uh so llm if you're talking about the llm so the base start from the Transformer itself just go to the Transformer using the python okay at least if you are able to code 50% then it is more than enough because in the industry you won't write the code from scratch until you are not going to build up your own libraries right uh if you are creating a simple PC or maybe in the uh highend project also you are using the buil-in library only like hugging phas or the buil-in apis right or other like a framework like py toou and all so no need to like write the code from scratch but in case if you want to build your own Library maybe on top of your custom architecture then definitely that level of coding is will be required great so depends depends on multifactors not the single one so it depends on the company and all so like for which company you are for for which like for which position and for which company you given the interview got it okay so let's start with the interview question so here uh list down the different different CNN architecture I think you can question you can answer for this type of questions list down the different different architecture of the uh list down where is the list down here is a list down the different different architecture of the CNN uh just talk about three to four more architecture and uh try to talk about your architecture and always go with the example right don't discuss the theory because no one interested that try to connect it with the real time Real Time images realtime scenarios then only people will listen to you otherwise they will reject to you so don't uh like even explain the theory of each and everything try to connect with the example and if you're able to connect with the real time example that would be that would be great okay now list down the name of the object detection algorithm so I think you can answer for this also different different object detction algorithm I told you now don't worry if you're worrying about the answers so definitely I will provide you the answers as well but yeah I have given you one more additional PDF there you will find out the answer uh like uh related to this question as well this type of question I have given you inside that PDF also but this is like a different PDF right which is we like uh there is like 150 to 200 question in a single PDF itself only the questions but yeah definitely you can prepare the answer of it and nowadays there are many resources uh available like uh you can take a help of the chat GPT jimy and all just for the quick understanding if you want to understand in depth definitely Google is there but don't worry if you want answer then definitely I will prepare one PDF related to that and I will provide it to you so directly you can learn from there as well but the main thing is what questions so yes you are getting the question and you can prepare in your own way as well after learning this much of fundamentals from me okay so what is the difference between object detection and the classification I think I clarify this thing also what is the detection difference so what is object detection localization plus classification and what is a classification the entire image we are passing through the convolution process I hope you are getting list major task we perform in CNN so what task we can perform by using the CNN tell me tell me tell me convolution is there pooling is there ra we can apply okay flattening we perform so different different like a task 1D convolution 2D convolution 3D convolution different different kind of convolution is also there getting my point here you can perform the padding and all right padding is also there means if you're losing like too much information so you can create one pad layer around the around your images around your raw images right so quantization of the model distillation of the model there are so many things quantization distillation is very important process nowadays for like making your model little uh like uh or light if you want lightweight model then distillation and uh like quation you need to perform where you are going to reduce the uh the weight the weight value related to your model okay so that is also very important part whenever you are going to build any realtime model and in the most of the research paper actually you will find out that right so definitely you can go and check segmentation segmentation I think I already told you what is a segmentation so the segmentation basically we are trying to create a outline around the object based on a pixel value we are not trying to find out the coordinate of the boxes what we are trying to do we are trying to go through the pixel to pixel and we are trying to create a boundary surrounding to the object okay that is called segmentation and in some of the domain actually you will um like find out that there is so many use cases Rel to the segmentation mainly in the medical domain itself right where if I want to detect something let's say if I want to detect some lump any lump okay in my like in the entire body right in the entire human body so definitely over the object detection I will perform the SE I will prefer the segmentation because precisely I can get that particular lump right or uh because here segmentation will perform well but simply if I want to detect the human in the images right or the like if I want to detect the multiple object in the like images I will go with the object detection there I I don't want to precise information I just want to detect that so yes definitely object detction will work for me right there are so many use cases like uh I think uh Rel to the traffic itself so I think you have seen the cameras which is deploying nowadays and automatically and automatically you are getting the like Chalan and all right e Chalan so how they are detected they they deployed the camera along with this algorithm so like they are focusing on you if your face is without helmet then definitely it will detect and it will classify that no helmet and uh if it is detecting with helmet then it will tell you with helmet so in the back end actually they have designed the system in such a way if you are you are getting detected no helmet immediately it will send the information to the like Auto it's automated system and the chalen will be generated right so this type of like this type of use cases application definitely you can find out and it is having a variety of the use cases so the in the entire drone industry right so the automated drone and the the uavs right or AAL vehicle or the vehicle everywhere they are using the technology okay this this particular technology detection and the segmentation and it is very useful and in many cases there is like many more like use cases related to that okay and even we can deploy this r model in the mobile system also so I think you are like you are opening your camera and it is getting detected your face right and it is telling you like what is your age and all so yes this type of functionality basically it's very common nowadays and how it is possible because of this CNN so definitely you should be learned the like uh the different different use cases regarding the different different domains right and that is very much important now the next one so which algorithm for the PCB right and here pre-train model or okay and here are talking about the transfer learning and here explain me your CNN network uh when it will fail so failing example and where we can use the RNN one so which GPU you use being to train your object detection model right GPU which GPU like they are directly talking about the hardware and all so definitely guys definitely so these are the very important thing if you are going for any interview you should like always prepare these type of question now what do you understand by FS explain 2D 3D convolution right what do you understand by normalization batch normalization okay handwriting detection how you can do the hand detection of the handwriting right so how you are extracting the data from the images and then basically you are going to classify that so this is all the the question now soft Max function disadvantage of RNN okay lstm so from here the RNN lstm is getting started so till here actually there was the CNN the fundamental question of the CNN but what I was thinking that this question is not enough so here I prepared a few more question on top of this computer vision right so here you can see so there are a couple of more question with respect to the computer vision and now let me discuss those question as well and here are around you will find out 50 to 60 question let me give you the glimpse of that and I will keep it in my itself so directly you can download it from there only so here what I'm seeing so just wait let me uh show you so it is basically based on the computer vision itself so what do you mean by the computer vision I think you know about it what do you what are some of the machine learning algorithm available in open CV right here you will find out how many type of image filters are in open CV Right image filter means filtering on top of the images and open CV is a library generally we are dealing with the images and all right it is having extensive functionality this open civil Library if you worked with a like Vis be system or maybe if you have work with the CNN or object detction object segmentation or any like use case related to the computer vision definitely you must have gone with this computer vision library because it is having so many functionality with respect to the images right several type of pre-processing we can perform by using this computer CV open CV library right now here you can see what is a pH recognization algorithm so there are a couple of algorithm basically which has been designed specifically for the face recognization now here you can see what are the algorithm or programming uh framework for supporting this computer vision what is the color model right so RGB I told you what is a dynamic range right what is a digital image what are Thea erosion and uh dilation right now which method is used to read the image in open CB now many more question basically we will find out to the the computer vision explain the segmentation to two type of three type of segmentation is semantic instance and panoptic okay you might get the Anchor Box so here explain what the match band effect is what are the drawbacks of the VG net so here I think there are so many question and definitely you can prepare it and which is directly related to the computer vision only right so yes guys uh and along with that I would suggest you please try to create some project as well because project is plays a very important role right if you're not preparing a project if you're just like reading the interview question and if you're going an interview definitely there is no use of that so always create some project and talk based on that particular project now from U Back side so here let me show you some more question how would you encode a categorical variable with thousand of distance value right so what is it mean to Inception architecture of CNN now here popular computer vision libraries how many type of image filter in open CV now here you can see rgba model of color representation what is a representing now which is the following is example of lowlevel filtering a lowlevel feature okay now here which is the following method is used to model fitting method for age detection so definitely guys just go and check with that and uh you will find out like a lots of things and your understanding will be much clearer and always try to uh work with the project and there use the uh different different model just do the experiment on top of the based on the different different model and try to make a conclusion that why this is working why this was not working right if you're preparing in such a way that definitely your chances is very high that you can crack the interview related to this specific positions so tell me if this thing is clear then please do let me know so I given you the glimpse of the computer vision as well so tell me guys is it clear then I will come to this RNN part and I will give you the complete uh detail of this RNN and then we try to discuss few question and then I will jump to the generative Ai and the project related question so if it is clear then can I get the quick thumbs up is it required to learn unsupervised learning unsupervised learning yes so you should learn the unsupervised learning also because with that basically what you can do you can uh find the you can find the pattern from the data right you can uh categorize your data right you can categorize your data based on a similar pattern so you can put it you can use it this unsupervised learning on any sort of a on any sort of a data right so yes unsupervised learning is also very important now where it is using specifically so that you will get to know after seeing the architecture only got it so now this thing is clear so let's discuss about the RNN based question so here is a fundamental question uh about the precurrent neural network now what I can do so first of all I can explain you the RN and uh then I will discuss the then I will discuss those particular interview question so the session is going fine guys yes or no tell me are you getting everything from the session itself yes can I get quick yes or no tell me and interview question wise don't worry so I will provide you all the interview question uh in my GitHub as well as from the resource section also you can download it now let's start with the rnl that is recurrent neural network okay so let's start with the recurrent neural network so we have one more branch of the deep learning that is what that is the recurrent neural network okay so what I can do here I can write simply I can write RNN now this RNN actually see in the RNN what we do uh we pass the sequence of information because why so let's say this is what this is my architecture of the Ann right this one this is the architecture of my artic icial neural network now what I'm doing I'm connecting it with each and every node this one like this and this is what this is my final output okay and here is what here is my input so this is the architecture of what this is architecture of the simple Ann now see in RNN actually what we do so here let's say we are passing a data so what I am passing I am passing the data over here now how we pass the data see so for the noral data right the data basically which is available in the table of format so simply we have this R artificial little Network so if we are talking about the images so if we want to press process the images yes so this is what this is my image so simply we have this uh Ann right so directly we can pass to theend but before the Ann basically some pre-processing is required pre-processing is required at the end everything is going to be connect to the Ann itself now here let's say if we are talking about the next type of different type of data let's say text right text so in the text actually the sequence matter sequence means what sequence like uh let's say here I'm writing mira mira n Mira now sunny high okay so here you can see the sunny actually it is directly connected to this miror right because I'm introducing myself so this sunny actually is connecting to myself if I'm not able to like maintain the sequence so definitely the sunny is not like there won't be any use of Sunny because the sunny will be unknown in that case the sunny actually he will be unknown in this in this case why because if it is not connecting to my my if it is not connecting with any person or to myself right so definitely the sun is unknown we won't be able to capture the context of the data right context of the data so that's why for solving this situation so where I can maintain the context of the data so we have introduce one more type of architecture that is called rnl recurrent neural network so here this R actually it is representing to the recurrency in the data recurrency in the data so recurrency how we can maintain the recurrency so recurrency actually I can maintain let me show you here I can write uh re currency okay so here you will find out that so let's say here I'm passing the input so the input is going from this input layer now here I have assign the weights right so this is what this is my weight there now this weight actually is going to be processed from the hidden layer from this one and then after that I will again pass to my hidden layer and here how I processing the data here basically I'm processing the data word by word this one so either I can there are so many techniques for converting this word into the vector right so either I can use the 1010 coding which is a very uh old technique and no one is using nowadays because it create like too much sparity the second technique is I can perform the word to back right word to back I can perform the word I'm reading now other than this we have other Advanced technique like I can perform the word encoding and all right word b encoding Elmo Elmo right we can perform the Elmo word based encoding okay and definitely we'll have to be encode this particular character why what is required because this model actually it's a mathematical model mathematics calculation is going on okay mathematical model and mathem macal calculation is going on let me write over here one more time so mathematical calculations is going on and definitely it won't be able to understand right they won't be able to understand this uh Hindi or English maybe right the high level languages so in that case we'll have to provide the number so because in the back end this number is going to be converted into the binary one so that my CPU can process that okay because CPU is only understand the binary system and everything is going to be converted into the binary form in the back end okay so I hope this is clear to all of you and uh this RNN architecture is also clear so word by basically it's going to process the data now here what I will do so here we have a different variance of the RNA so some Advanced name let me write here so RNN is there then basically we have lstm then we have Gru as well so Gru and this lstm actually it was a very famous one so lstm was there now here this RNN actually it is of Performing the sequence data right so we have we have different different type of sequence mapping right we have different different type of sequence mappings over here sequence mappings so sequence mapping guys here uh let me introduce few sequence mapping over here so let's say here what what I have so let's say this is what this is my model right so this model actually uh I can introduce you one more time so here let's say I have the model so uh here the different different sequence mapping let's say one to many one to many here many to one many to one okay many to one and here you will find out many to many many to many right many to many now if you want to look the architecture so definitely I will show you the architecture as well so we are passing only one output and we are getting a many output we are passing the many output we are getting only one output we are sorry we are passing one input we are getting many output we are passing many input we are getting only one output we are passing one uh many input and we are passing many output so it is based on an input and output now input and output right input and output now guys here there was a different different type of mapping technique now uh here actually what they did in uh 2014 actually there one Advanced mapping technique so the mapping technique was same right just many to many one advanc concept has been introduced the concept was the encoder and decoder encoder and decoder so here actually we were passing the data okay from here uh means we were encoding the data in terms of context vector and we're passing to the decoder so this encod and decoder is actually nothing it's a sequence to sequence mapping itself and it was the many to many mapping many to many mapping right and there itself they were using the lstm RNN or Gru architecture this one either RNN lstm and Gru right this one and then inside this encoder decoder architecture one more thing has been introduced that was the attention one attention right attention so attention actually so here what they were doing inside the attention so every hidden state of this model OKAY every hidden state of this encoder it is going to be correct every hidden state of this decoder right so like this this one so every hidden state of this encoder is going to be connect every hidden state of the decoder so that was the attention model okay and after that one more paper came so that was a self attention and the entire uh generative AI basically the generative model which you can see right large language model it is based on this it is based on this self attention paper only so entire large language based model LM model it is completely based on this self attention paper right the paper name is attention all you need attention all you need attention all you need right means you just require the self attention nothing else right for capturing the data now if you want to look the architecture and all so I'm not going to teach you anything over here actually why because everything is there in my uh playlist in my playlist of the gener that is very uh like uh uh successful one also let me show you that part playlist and uh if you want to learn about the generative Ai and the complete timeline from RN into lstm right one more thing I tried to forget I explain you uh I forgot to explain you there that was a bidirectional bidirection right so we can process the information in both direction that is called B directional so maybe like you have heard about the bidirectional by Direction RN so instead of processing the information from single Direction you can process the information in both Direction in this from left to right and right to left that is also possible uh that is also possible right so yes bidirectional HDM bidirectional RNN and even but also the transform based architecture it was processing the information in both direction right encoder one encoder part it was processing the information in both Direction because we want to understand the context of the word okay from both direction right we have couple of example I will come to that but yeah here uh first of all let me show you the playlist so that you can learn the generative vi from there and uh it's a very uh like refined playlist on top of the generative a so if you will go and check with the uh ion YouTube channel now here just uh search about I neuron so here you will search about the I neuron on and then go inside the playlist and inside the playlist guys here you will find out the generative AI playlist so Foundation of generative AI see guys this playlist actually it's amazing one now just look into the first uh first lecture of it introduction of generative AI for the community so just look into the views 45k views and here you'll find out the complete detail of the gentic AI that how it evolved from NLP to the generative a right how we reach to the NLP to the generative a everything basically you will find out inside this session now if you want to learn more about the generative a then definitely this playlist is for you if you want to learn as a uh in the free course right so yes definitely you can enroll inside the course as well and you will get the access of the dashboard but let's say if you want to learn in a specific course definitely we have launched one uh paid course as well related to the generative AI itself you can go and check with website mastering generi with L CH Lama index and whatever framework is there in the market and I think we have spent one month as well uh in the live class and there I discuss just the foundation of the generative a that cleaning encoding embedding or transfer learning pre uh fine tuning and all using hugging face Transformer and all right now coming to the main in that particular batch now I'm going to be start with the main topics so please go and check with that batch also mastering gen with llm and lenen you will get it over the Inon website itself so I hope this thing is pretty much Clear Genera VI and all now let's back to the interview questions here so first of all guys what I can do I can show you the fundamental interview question of the rnl STM and all and then I will come to the generative AI based interview question and then finally project related interview question and then tomorrow resume and the profile building Okay so here you can see what is the disadvantage of the RNN so what was the disadvantage of the RNN list down at least five RNN so different different variants of the RNN that how many type of uh like uh sequence to sequence mapping you have related to the RNN right and by directional RM lstm or Gru right it it's just a modification on top of the RN itself the core unit of the of the processing is nothing it's the RN itself getting my point tell me so how uh are you getting it guys yes or no tell me are you able to get it uh I think whatever I'm teaching to all of you maybe you are getting each and everything now now tell me guys can I get quick thumbs up or something uh and uh please like the session also if you are able to understand each and everything and if uh like this interview question you are thinking that definitely it might help to us uh for the interview then you should give me a one like at least so tell me until I can drink some water and then I will start great so I hope you can see the various question based on this RNN lstm and all just try to read it and let's try to understand so here five type of RNN means different different type of sequence to sequence mapping different different type of like variant of the RNN and all now here explain architecture diagram of lstm so yes that is also very important part explain architecture diagram by directional lstm okay now here explain architecture diagram ST lstm lstm means so there we have a multiple hid layer getting my point by using the single hid layer we are not processing the information we are processing the information by using the multiple hidden layer and that is called stacked RN along with the advantage and disadvantage you have to discuss this thing guys okay always you need to discuss along with the examples example role we're just talking about the theory no one will believe on you until you are not giving the example until you're not convincing to the people right so Advantage disadvantage very important Now understand what do you understand by TF IDF TF IDF is nothing it's a way to convert the data into numbers that's it term frequency and inverse document frequency so term frequency multiply with inverse do frequency and that is called tfidf and tfidf is nothing it's just a like a text representation in terms of numbers so definitely you can create a vector by using this TF IDF now word two actually it's a neural network based technique so we are generating an embedding by using the neural network getting my point so between the last hidden layer and the output layer whatever whatever weights you will get right so you are trying to represent your like your words in terms of that weights okay so that is nothing that is called word iding word iding you are going to convert into iding specifically uh what to Vector what to embedding why they are using Vector embedding this word because uh they have introduced one New Concept in 2012 itself I think 2013 they have introduced this word to back and was a research of the Google itself tfidf uh like people this Google was using okay what to bag also invented by the Google itself you can check with the research paper of it you will get the complete detail that how they are using the neural network for converting word into the vectors now the next one is five vectorization technique so everything is around the numbers only right so how efficiently we are converting our data into the numbers so this vectorization is plays a very important role in self attention what we are doing tell me we are trying to relate every word with every word with other words that is a self attention now means how efficiently we are going to generate an embedding how efficiently we are going to uh how efficiently we are going to retain the d dynamical uh Dynamic context of the sentence that is a self attention and everything is about the embedding only getting my point yes or no so vectorization technique means different different vectorization technique like tfidf okay uh this is the count base frequency base now we have other techniques so other techniques like what to back Almo fast Tex board based encoding and openi based jimy based right so these are different different eding technique and they in the advanced one like jimy openi and the board based encoding they are using the self attention and then they are passing it through the like neural network and finally it is giving the vectors okay vectors from the Ws getting my point so here uh this is also very important different different word vectorization technique you should definitely learn now some uh metrics or this metrix factorization based Technique One technique glob okay glob Vector you can learn that is also very efficient one right glove Vector now what is the differen between RNN and encoder decoder so RNN encoder decoder so RNN is what RN is architecture which is simply processing the information right so in a different different manner we can map the information okay one to many many to one and many to many but many to many information which we are mapping by using the RNN so there we are not able to map the as synchronized data which is having a so let's say we are going to perform the machine translation if we are going to solve the machine translation so left hand side we have a different length and right hand side basically the translation which which I'm getting that is having a different length so it is not going to match so there this RNN is going to be fail so that's why we have one more sequence to sequence mapping that is encod and decoder so like this basically you can make a differences between this encode and decoder attention mechanism what is the meaning of attention mechanism so attention mechanism is nothing so we are trying to con like connect every hidden state with uh with a decoder with the every hidden state of the decoder right every hidden state of the encoder connecting every hidden state of the decoder that's it that's the meaning of the attention and how we are going to connect it by using the neural network because with that only we can automate the process otherwise manually I will have to find it out manually I will have to generate a vector and that is not possible for the human for the bigger data set okay so neural network some function will be required a mathematics fun mathematical function okay mathematical function so definitely we'll have to use a neural network over there now here attention all you need and what do you understand about the multi attention and all this is a question and related to the Transformer basically so yes this fundamental question you should always learn and apart from that you can go and check with my personal YouTube channel also where I'm uploading the content on top of the generative VA if you interested then where you will get the detail and the like extensive road map of the U like gener VI and all right so here I YouTube channel so I have already introduced playlist if you want to learn the generative AI along with the implementation applied generative AI definitely you should check with this playlist it is for you only okay if you are like looking into the if you want to install in the field of generative AI you will get the foundation of the generative AI here and then you can go in depth according to your uh like your wish right so I hope this thing is clear the fundamental of NLP is clear guys what kind of question they might ask you and always go with the project guys always try to believe on the implementation always try to prepare the projects if everything is perfect then can I get quickier so that I can discuss few generative AI related interview question also tell me guys fast if everything is perfect then can I get quick yes or no tell me tell me tell me guys fast so that I can proceed great now what I can do so first of all I can explain you that okay so what I can do here guys just wait so I can explain you from the uh document itself right and don't worry this all the interview questions and all you will find out inside the resource section of the uh resource section of the ion Dash so here guys the first interview question is what is a generative a I think you know what is a generative a which is capable to generate something okay so Generation generation how it is going to generate tell me so based on the training right based on the training it is having the capability so that it can understand about the data right so that is what that is a generative AI now uh the next one is what how does gen VI works so in a similar fashion right so in a similar fashion actually gen work so I think you have seen the classical uh I think you have seen the pipeline of the classical machine learning and the Deep learning yes or no so in the classical machine learning and deep learning what we are doing so first we are going to collect the data let's say here I can explain you so first what we are doing we are going to collect a data right so first we are going to collect a data then what we are doing after that we are doing analysis of on top of that analysis on top of it then what we are doing we are performing the pre-processing okay preprocessing then here after the pre-processing tell me pre-processing then we are going to create a model and this model actually we are creating from the scratch and then finally the evaluation of the model so here we were having the evaluation stage now if we are talking about the generative guys so there also see uh there also we have the data will be required actually uh generative a uh see it works on uh this particular technique the transfer learning transfer learning transfer learning right so if you want to create any sort of application okay so based on the G model let's say if we are talking about the llm large language model so we always perform the transfer learning on top of it now in the transfer learning actually what we do see the model already has been trained the model already has been trained right model already has been trained now here for the training actually they have followed the same pipeline this particular pipeline where they are going to collect the data analysis the data pre-process the data then it is passing to the model now this model actually it is going to train in such a way this model actually is going to train in such a way so that it can understand the it can understand the pattern of the data it can understand the pattern of the language getting my point here right so this Lang this this model actually is going to be trained in such a way it can understand the pattern now here there are several technique which is involved inside the model training right so what I can do just wait let me explain you over the board itself so I think that would be a great experience for all of you so what I'm doing let's say if we are talking about the machine learning pipeline so there we have a several stages so first stage is called Data injection so where actually what we are doing we are going to ingest the data the second stage is called anal is of the data right analysis of the data so analysis of that data the third stage is called a pre-processing preprocessing okay and the third one is called sorry fourth one is called Model creation model creation and the fifth one is called fifth one is called evaluation right so what we do in the class classical ML on the classical DL so first we perform the training right first we perform the training okay in the sorry the classical here I'm going to write the classical ml so here in the classical ml what we do classical ml or DL so first we perform the training by using these particular step right what we do guys tell me we perform the training and then then prediction right then what then the prediction okay now in the prediction actually we just required the data so here let's say we required the data so this is by data we are passing it to the model this particular model The Strain model okay this particular model and then we are generating a prediction based on the use case right so here is what here is my model this one there is my model right now this is what this is of prediction now this type of uh process is called the discriminative process right discriminative process discriminative process means here based on the pattern based on the data itself is going to generate the output means it's a use case specific use case specific use case specific I can give you very good example of this use case specific let's say here by using this model uh you have trained this model on top of one data let's say it's a classification problem it's a classification problem right it's a classification problem now here you have done you have trained this particular model on top of this problem statement so where actually you want to identify that whether the person is having cancer or not right cancer or not cancer not cancer right so you want to identify this thing by using this particular model you have trained your model on top of this particular data right but let's say instead of passing the feature related to this data whether the person is having cancer or not you are passing a different data let's say there you are trying to find out the person is having the person is having diabetes or not diabetes or not diabetes or not diabetes so in this in this case guys in this case you'll find out diabetes or not diabetes in this case you will find out this model is going to be filled even this model won't be able to understand the feature it will tell you the feature which you passing it is incorrect the feature actually which you are passing the column which you are passing it's incorrect so it's a problem specific or use case specific model and this type of model it is called discriminative model which is called the discriminative model right and in the classical machine learning and DL learning deep learning actually we are creating this uh user uh this use case specific model or discri discriminative model I hope you are getting it I hope uh you are able to understand what what I'm trying to uh what I'm trying to explain yes or no guys tell me so this discriptive model is clear to all of you yes or no can I get quick yes quick no in the chat yes use case a specific model and I think you can see the clear-cut architecture over here of this use case specific model right now the next type of model the different type of model here what I can do I can take the generative model right generative model now actually whenever we are talking about the generative model okay generative model so generative model actually see this is also having two phases so the first phase actually which is called training training and the second phase is called inferencing or the prediction right inferencing or the prediction inferencing or the prediction or here I can write the prediction now guys see uh here if we are talking about the training so here also the same step is involved so first actually we going to collect the data now here we are going to analysis the data because without analysis we cannot perform the pre-processing here we always use the statistic then the pre-processing is involved pre-processing is involved and then the model building model building right then the model building and then evaluation of the model common steps right common steps but here guys see whenever we are talking about the generative model so the first thing let's try to make a differences from the classical model and this generative model right classical model and the generative model so here guys see first thing is what the data will be very much H here actually the data will be very much huge huge data okay and and here guys we are taking a huge data definitely analysis pre-processing will be involved whatever pre-processing analysis based on the data then the model building guys see here actually we are going to train a model on top of the hug data and here we are going to train the model in such a way so it can understand the generalized pattern of the data generalized pattern of the data generalized pattern of the data means it's not a use case specific it's not a use case specific data sorry it's not a use case specific model so it's not a use casee a specific model it's not like that right it's a generalized model it's a generalized model right now here the training process will be also it will be also like different so here actually there in the classical mod so either we performing the supervis or unsupervised right supervis or unsupervised right either we are performing the supervis or unsupervised but here actually we are performing both let me tell you how right so here guys see first what we do while we are training the model OKAY in the generative model let's say I'm taking example of large language model large language model first of all let's TR to understand the language model so language model actually it always teach the language to the model OKAY language to the model OKAY the model basically which understand about the language now this model actually this model it's a very HED model okay so here I can write large here I can write large right so this model actually it understand the language and it is it is Stained on the huge amount of data so that is what there is a large language model right now here if we are going to train any large language model llm model right so llm model now here it always use Transformer as a base architecture Transformer as a base architecture Transformer as a base architecture now Transformer basically it's architecture which is using the neural network along with the self potention right neural network along with the self potention now this architecture actually it is required use amount of data use amount of data so for understanding the pattern right so for understanding the pattern from the data actually so we always go with the language modeling right language modeling and here actually we perform the unsupervised language modeling unsupervised language modeling unsupervised language modeling and then on top of it we are going to perform the supervise fine tuning and that is Task specific supervise fine tuning so this this one UNS supervis this one unsupervised language modeling and the supervised fine tuning basically right so this is here basically we perform the uh like here like this we always perform the model training right by using this supervised unsupervised language modeling and the supervis fine tuning with respect to what with respect to this large language model which is using this Transformer concept okay that is nothing that is a neural network along with a self potention for understanding the data in a better way so I hope you're getting that how this model is different now whenever we have to do the inferencing guys right so this model actually now it's a general model it understand everything it understand that how to summarize the language how to generate the next sentence it understand that how we can fill the gap between the sentence right how we can classify the sentence right you know about the GPT chat GPD it can do everything it can do everything how because it it understand about the language it understand about the language how because we have we took the huge amount of data and we have trained it in such a way in this particular way and that's why here whatever prompt we are passing whatever input we are passing right and based on that it is going to be generate the output the specific output based on a based on a input but here in the classical one okay here in the classical one basically what we are doing so we passing the feature a specific feature to the model a specific feature to the model and according to that basically we are generating the output according to that basically we were generating the output and that is a that is a like main differences now coming to the interview questions and let me show you the interview question if this major difference is clear to all of you related to the generative model and this classical model tell me guys is it clear tell me each and everything basically I try to explain you here in a simplest fashion if you're following to me definitely you will be able to understand if you are finding out this writing is not correct but if you will go with my words definitely will get that also I hope this is perfect now let's try to move to the system again so here is a system now what I can do I can go through with the interview question one by one now how does gener works I think I given you the entire pipeline now what are the top application of the generative AI so top application basically natural language based use cases NLP based use cases most of the time you'll find out that only like uh text classification text generation or uh machine translation okay translation will be there generation will be there classification will be there summarization will be there right so these are the like task we can perform even we can uh perform the like domain specific task also so but for that actually we'll have to fine tune the model directly we cannot use the transfer learning we can create the rack system okay so they are actually fine tuning won't be required again training will will not be required directly we can fetch the data from the r from the database itself and according to that we can enhance the output right so here are like couple of use cases you can go through with that can you explain the difference between gen and the descriptive a so that is a same thing basically I try to explain you difference between generative VI and the descriptive VI what are the some popular generative models so here GPT BD dally okay Style gang or this Vector quantize variational Auto encoder so this is the generative model and like nowadays we have so many so many means so many right and I'm recommending you this particular playlist that definitely you can go through with it and you can learn it now you can see couple of more question related to generative AI so how generative AI gains works it's a fundamental of the generation before the llm model actually everyone was using the gains for generating images right limitations and all everything basically you will find out over here so what we can do we can discuss rest of the question in tomorrow session because it is already five and then after that I will come to the interview sorry I will come to the resume preparation and the resume building sorry uh resume preparation and the profile building okay so I hope guys you like this session now I have so many questions here so I want to discuss uh like I want to discuss all of them but uh not like this okay step by step one by one at least I will take half an hour because it is important and along with that along with this one basically I have the like project related question also so both we try to cover up within a half an hour only now after that uh basically we can go with the resume and the profile building and all so tell me everything is perfect everything is clear did you understand the differences and did you understand the fundamental of each and every if yes then please give me quick yes or no and this interview question basically will help you to uh like crack the interview believe me guys if you're going for the interview then uh definitely you can read this interview question and uh you can go for that you can go for the interview now uh one more thing guys let me show you one uh production ready batch because over there actually we are discussing uh like a different different project related to the different different uh is St so here mlops production ready data science project yesterday also I introduced this one now if you want to enroll to this production ready data science project definitely you can enroll recently we have started this batch and in this one so we are discussing the different different like a projects okay related to the different different domains so machine learning CNN or computer vision NLP regarding any sort of a domain right so related any sort of a technology you can find out the project over here and the detailed description of this project because Theory you can learn from anywhere but when it's uh comes to the implementation at that time you find out the difficulty that okay so then you will search like various YouTube channel then you will search like various GitHub repository and at the end like you are just copying the project and you won't able to get that how to design the project from scratch and and that is a problem of majority of student okay whoever is working in indust even they are also not they are even they are also not find out the specific pipeline of the implementation related to this ML and DL project even the generative a also so guys if you want to steam like that pipeline if you want to create that pipeline automat uh if you want to create that particular pipeline by yourself that definitely you should join this uh batch and there you can Implement everything from is scratch right so me boy both are teaching like a different different projects over here you can see the mentors even Chris is also there so you can get the complete Mentor detail over here on top of this homepage so I hope this thing is clear to all of you this is a kid level is not going to ask you this interview for me it was very technical the nature yes you can discuss like what type of question someone has ask you but yes fundamental start from here in like definitely there will be Advanced question also that's why I'm telling you now just focus on this one Advanced question wise let's say if I'm saying Vector indexing with what type of vector indexing you have used while you are like uh passing the data into the vector database now you don't know the embedding that how you will be able to understand the vector embedding so should I directly start from there or should I first tell you the embedding uh the fundamental of the iding tell me if I will explain you the fundamental of iding then definitely you can directly relate with the indexing also right yes or no I understand people will ask you the advanced question but they will begin from here itself it will begin from here itself now here if I'm asking you okay so how you are performing the positional encoding how positional encoding is helping you for capturing the dynamic context tell me if you don't know about the self attention mechanism if you don't know about the key G pair in that case how you can do that how you can perform the positional encoding on top of that so that architecture should be clear first and then only you can go for the next level okay so yes first beginner then intermediate and then Advanced and how it will come it will come from here itself got it great so fine I think uh now we can conclude the session it is already five and uh definitely we have uh so we have one more session on on the YouTube channel if you if you are willing to join that like C++ with DS definitely you can do that so fine thank you guys thank you for joining this session don't miss tomorrow session because it's going to be a last session of this community series of data science interview vision and the question basically from basic to advance only fine so thank you bye-bye take care guys have a great Daye tomorrow we'll connect on same time at 3 p.m. IST 3: to 5:00 p.m. so until thank you bye-bye take care

Channel: iNeuron Intelligence

Views: 3,594

Keywords: machine learning interview questions, machine learning interview questions and answers, machine learning interview, data science interview questions, machine learning, machine learning engineer, machine learning course, data science, machine learning engineer interview, machine learning tutorial, ml interview, ml engineer interview, ai interview questions, data scientist career, programming interview, python interview questions javatpoint, python programming, ineuron

Id: F1lsFTpsQLI

Length: 125min 45sec (7545 seconds)

Published: Tue Feb 27 2024