Live Virtual Mock Interview For Data Science Role

Key Takeaways

so this okay guys we are live we'll just wait for some time so that everybody is available and i've given the link so don't show in the chat you can check it out i will um we'll just wait for two minutes guys so that everybody joins and then we will start okay so 20 people have joined many people have joined now so welcome everyone as announced there before yesterday that uh we will probably be taking a live virtual mock interview sessions for this entire month that is the plan so the first person as usual uh ravi tanwar is with us and if you don't know about ravi uh we'll just uh you know he's also a kaggle uh two times expert uh and uh we'll just know more about him so welcome ravi once again uh for this virtual live virtual mock interview i hope you learn from this interview a lot and uh you know you have fun also and probably we also have with us sudanshu is the ceo five neuron obviously you know him uh as we have been taking a lot of interview sessions so ravi uh should we start anything that you want to say before this before we proceed no i just i think i'm ready but time will tell how much so okay now uh just uh just to make you know ravi this will probably go for 45 minutes to one hour and uh the plan will be that the entire interview session will be divided into parts so first 15 minutes probably python will be there then we may go into stats then just by seeing your resume we may proceed with other ml or dl topics okay so i hope sudanshi are fine with it right yes so let's start without wasting any time so again ravi welcome to this interview before going ahead uh the most common uh question is that uh tell me all about yourself yeah yeah so hi guys so ravi here and uh i've been uh so going back to a little bit about my degree i finished my graduation from punjabi community college in 2019 and upon graduation i i started working at a startup and upon uh working for uh for some time i got enrolled into a data science career track by springboard and while continuing that i got a job as a business analyst at my present company and a few months into the business analyst role i was i was given we can say a promoted so thing to a to a more data science and data science kind of role so over here i'm doing uh data science plus nlp kind of work and that's about it and apart from this i've also collaborated with a few other youtubers in this data science space and i've also i also used to participate on kaggle not so much anymore but i want to continue that and apart from all these things uh i'm a classic rock and old music fan and you can find me on linkedin i'm pretty active there and that's about it okay okay so uh fine i can see you like a resume ravi so as of now you are working with a company with a startup and before that so i can see there are like a kind of a series of startups with which you are working and i can see your github uh as well so there are in total 57 i think repositories right so for sure i'll have a look into that write your github repository up so i think some of the github repository is like good i'm just going through it like in last five minutes so i have seen uh many things that you have mentioned over here so again you have uh mentioned about the lead code solutions as well right so it looks like like you are pretty good with coding stuff right and uh yeah so as you are working into our data science so far sure you must be very good with that one okay so i'll find ravi so what what we can do is like uh let's try to start these things with a python first of all and then probably we can try to move into a machine learning part a little bit and then we'll try to explore a little bit of deep learning and then nlp as you have mentioned so you are working into this nlp part right so for sure we can we can ask some questions from nlp parts so it's fine right really sure sure okay fine so uh fine uh ravi so i have one data set with me and i'm assuming that you have joined with your system so can you please uh like uh check your chat boxes and i just given you a link of one data set so it's a very common data set opens data set right so like uh i have to give you some tasks based on that data set so if you can access it and if you can share your screen and you can try to open up any ide so whatever whatever you like so you can just try to open it up and whatever language you would like to use so i don't have any kind of issue with the language so you can try to use java python art programming or maybe like a dotnet or maybe any any other languages that you know right so i'm completely fine with the language so there is no restriction with the language part but yeah so we have to complete some tasks okay fine that's great so this is our data set so can you please load it yeah just just load this data so is my screen visible right yes visible yeah cool okay okay so fine so here you can see a multiple columns let's suppose like a year column is their month column is their day column their gender column is there right so first of all a small python task i would like to give you that so let's suppose if i'm looking for all the data set right based on one simple filter condition so wherever i have a float data point or floating data set into a day column and wherever i have a string data set in a gender column right so i will be looking for an index of those data sets so is it was possible for you to give it to me so for example let's suppose there is an a column let's suppose there is a gender column right so let's suppose i'm talking about index 0 number 0 okay now so you have to like uh verify that okay fine so if day is equal to one and if gender is equal to string like let's suppose days day data type is equal to float because as you can see it's a floating point number which is measured right day data point is float and if gender data points are objects or string you can say so just try to filter out all those columns all those records all those rows okay so checking the data type we have for the day it's a float as expected and for the gender is object right yeah yeah but if you will try this behold it has it right so somewhere you will be able to find null values as well so i would like to add all the null values based on the data types okay so for uh so in day there are uh four eighteen values okay yes yes so i'm not looking for a kind of a record uh which is having a null values right and uh i'm looking for both a check so on gender as well as on day although inside our gender column we don't have any kind of null value but still i'm looking for a check over there so i have also picked the data for the audience also if they want to try okay so um let me understand the question like once more so we have these two columns right and this this column contains uh only the floats yeah right yes nodes along with null values and the general contains only the uh what do you say string or object the object strings or object yeah so what is uh to be like uh filtered out like in this game i would like to get all the data set wherever data type of day column is float and data type of like a gender column is object okay just just based on data type check okay um df dot d type you will have to write in a beginning so df dot d types is gonna give me the all of the data types right if dot d type inside that df.gender dot d type is equal to object just just hint i'm just giving you a hint it's okay uh we still have 15 minutes let him try it df dot d type it's okay it'll be just uh some tricky things are there just thinking that if you really also want to check the internet uh it's okay you can check it but just uh stop sharing the screen and check once or at least try to see stack overflow solutions you know no no let's let's just uh discuss without uh the internet yeah structure so so what i'm thinking is okay i check the uh data types for this columns and this data type should be either object or float so this is what i'm trying to think at or shall we use some kind of a for loop and then check it don't have any issue i don't have any issue with the solution so if you just feel free to like uh use whatever you want a for loop will be very very tedious task you know that is the reason why you have pandas data frames right yeah you use it i don't have any issue it's fine yeah yeah obviously for loops will be slow what is the question krish the question is guys he basically have to write a filter condition to pick up uh the object values and the integer values from this day column and probably uh the gender column right yeah so this is the question okay guys if anybody wants to enroll for the mock interview just see my community post over there i've given you the uh google google form you have to just fill up that file okay [Music] [Music] hmm okay so with this what i'm trying to do is um i'm trying to get the role where the genders data type is the this yeah so you can write in this way ef dot d ties bracket inside that df gender dot d type is equal equal to object it will work in that way df dot d types initially yeah so initially initially that is fine that is fine inside the back it's fine outside the bracket i'm talking about outside the bracket from where you have started yeah so there are df.d types yeah so df.d types and then is equal equal to object you can write in this way because see inside that so what we are trying to do so df.gender will try to give me a series out of that so we are trying to find the data type of that right so df.general series of d type is equal to nothing but it's a like a similar object it is going to give you right and we are trying to like uh check from this one so from which particular like a data frame so we are trying to check these things from our another series which is the df.d types dfr types itself will give you series you are trying to check with like a data frame that's a reason so it's giving you an issue it's not an issue you can you can just try to write uh like uh object over here just see the chat i've given one solution just try it out with that probably instead of oh right so you can write of yeah object object complete object no no df gender df of gender right hope you have yeah gf of gender d type is equal equal to object yeah execute i know just just remove key value error so d types right for single column it will be d type right for what is the issue you are getting i think we need to use the dot reset index here i think you have used reset indexes right ravi just have a look on to the solution that i have given in the chat once so that is just like a hint given to you so that you can try hmm okay uh i think sudanshu is this kind of answer that you required um yeah so like a df dot columns and then okay or okay you have written all conditions so yeah so maybe you can just try to replace this with the end condition that will be it yeah so uh and i think in this we'll use or only no no i was looking for like all the columns all the data set where i don't have a null values so i have to remove values right so i can't accept those rows which is having a null values okay like that does this also yeah sounds great yeah drop in here so uh fine so this was the first one uh this isn't so that's just one question ravi just go in the next line uh probably you have uh two strings just write str1 is equal to act act act act in strings yeah htr2 is equal to c80 just try to write a check or a function which will basically say whether this two strings can suppose in string one you have this many letters right so by using the same letter you can also create string two right you see this so can you just do this create a function which will check two strings and probably it will say you that whether it is an anagram or not so this anagram it is basically a anagram okay whatever letters are used in string one you can use the same letter to create the string too so i just want to try to create a function it's just like a competitive programming saying that i'm just trying to create a function wherein you give this two strings and probably you should give true if it is anagram okay i hope you understood the question yes and probably will try to test it from other things also okay in the same way it you should try to write a way so that it checks for the conditions also actually it's a bit cold over here so my hands are sort of it's okay it's okay take your time yeah so uh what how i'm trying to approach this is uh first of all let's convert everything to lowercase and then replace all the blanks right and then when we have uh done this pre-processing step we can i think then check if the counter of s1 is same as counter of s2 do you think it will work well suppose i have three other characters in str1 this will be giving you true okay just try to do this okay return counter s1 double equal to counter h2 yes execute it now give s1 as something else don't give the same thing yeah uh uh what should i write in this s1 uh just write anything like 80 right just write 80 80 okay now give it str1 and hdr2 just check what condition it will come whether it will come true or false okay it's coming false what about the correct scenario just go ahead with that so like what do you mean but like correct just write act just write act okay okay so this should return true right yeah now can you implement this counter thing by normal code yeah sure so can i do this in a separate function or so yeah it's okay do anyhow you want just implement that counter function so we have still two minutes so after that we'll switch to something else so after python cool so so counter is basically uh a dictionary right so i initialize this as a dictionary kind of a thing okay then for i n s one right i'm iterating over uh this uh string or list for that matter and then what i'm trying to check is so in this so there are a couple of ways to do this so so so let let me do the first one first so if i in c what this means is if the key is already already present in this dictionary right then what shall i do is i shall do c of i plus equal to one just increment the count okay and if it's not present then do i equal to one okay okay and let's um return this c and then let's check if this works or not so really okay and so b occurs one time i occurs two times and uh l occurs two times a occurs one time so i think this is what uh i think you have got the idea now so that should let's continue with the other next one yeah sure so ravi can you please go to the same data set the data set that you have loaded yeah okay so now as you can see so this is basically a birthday does it right so if you will try to see like uh that there is a bother given starting from 1969 to 2008 right so last record if you'll go and check so 2008 now uh so maybe like uh if if i have to predict that uh how many male and female right uh will be born in 20 2009 or maybe in 2010 right so what will be your approach in that way so if this is the like a problem that you have that okay fine so you have to do a prediction right so what will be your approach uh okay just in just a minute so as we can see starting from 1969 and 1970 all the way till 2003 2008 in between you will be able to see okay 2004 is here two thousand eight one below 1994 so i think value underscore count it has sorted the data based on the counts so group by as well as like uh here and uh okay so okay 2008 is on the top okay so if you have to predict that how many child will be born right so into our nine so what will be your approach just by looking into this data and the thing over here is like uh you are you are trying to print this value underscore count so balance count basically it's just trying to do group by account operation right right a particular year how many records you have this is what you are able to see based on the code that you have written right so we have a 24 records but it is not showing you what was the total number of birth in 2008 so maybe you can try to print that and you can check that one so maybe you will be able to get some trend okay so so let's group by the year right and then let's take the burst column right and then we want to do some right yeah so now you can see the trend right no i think we should do count or some so you can try to you have done group by and then yeah and fine then you are trying to select this birth column and then summation of this one so it is going to give you total number of birth in a particular year so i think this is this is fine oh okay okay i thought this data set was small so that's why i was looking at how much how many this is such a high number of births okay okay yeah cool okay so so given this data uh we have to sort of check how can we predict the number of babies for uh 2009 right or maybe for next two years if i'll ask you right so give me a stats birth so how i will try to proceed with this one and again so if i will ask you that okay fine so i'm just looking for a male or female right so how many a male will be born or maybe how many females you want so how we can try to proceed with this data so if we take if we only look at the births right so um so this is somewhat of an upward trend but what i'm thinking is uh uh we can map this into some sort of a regression problem right okay and given the year and i mean what i'm trying to think is can we treat the ear as an integer or do we have to like do some kind of an if you are trying to treat here as an integer so for sure uh there will be uh effect on your final prediction right because individual will be having values may affect your final prediction right so my situation will not to treat ear as a number here as an integer over there this will be my suggestion but yeah try to tell me your approach yeah so that's why i was the i was uh kind of not sure whether i should treat this as an integer so uh what is coming to my mind is that uh maybe we can do some kind of uh uh it's already in some kind of a label encoded form right this year thing because there are multiple instances of the same year so either do some kind of a label encoding kind of a thing and then try to form this into uh and try to phone this into some kind of a regression problem i was also like thinking in form of lstm but uh i'm not really sure about that because in lstm we have uh sort of a sequential data right so we can treat uh so we can treat the training set as the sequences of a few years and then sort of use that sequence to uh predict uh another sequence say okay so what will be your final approach like are you going to apply regulation or are you going to like go and pick and choose for lstm based approach or maybe is there any other approach that you can think of this second so it's not a small data you have like a 15 000 records and i think it's sufficient enough to build at least a basic model not not a stable one but yeah maybe a basic one right right so as of now i think uh only two approaches uh come to my mind if i have to treat this into it doesn't make sense treating like according to me i think it's feasible if we treat this into a regression problem okay but don't you think that that if you are going to treat these things as a regression problem and if you are going to consider ear as a numeric values right so in that case is going to affect your label data it is going to affect your like labels in that case yeah right right so that's why i was sort of thinking if we could convert these years into some kind of a label format maybe uh maybe instead of numbers we convert these into some other but again trying to do like a label encoding right let's suppose we're creating a case so where i get i can go and convert this things into a label encoding situation change much for me right because let's suppose if i'm going to start from innovation 0 1 2 3. again situation will be same maybe i will be able to reduce our variances right possible but uh again situation is not going to change over here what is the kind of solutions which i am supposed to build what is the approach i'm just looking for approach nothing else right right so going by this rational the regression is the regression won't be a desirable one because the right right because these are both are numeric features right so even if i do and some kind of label encoding it's gonna in turn convert this into a numeric one so that uh that won't be uh relationship between x and y so so can i can i think of like uh using some time series based model or maybe let's suppose lstm itself so is it possible to use lstm over here so if i'm going to use lstm uh what what i should do over here or maybe uh i can try to use some time series based model arima model ceremony model arma model so any any models i can try to use so any suggestions for that so i uh so when the regression approach sort of uh wasn't working out so i think that's when um i thought like yeah if this is not working out so in that case is another feasible approach is obviously using converting this into some sort of a sequential data problem right so whenever like situ sequential data is the case uh obviously as you said we can use some sort of uh either a time series approach or or basically uh treat this into uh into training data of the various sequences various uh length sequences right uh like uh like grouping the year 69 to 71 and then 70 to 73 this sort of thing so that we have a sequential training data and then fit the list em on that training data so that it can predict uh for uh upcoming three years maybe yeah but so lstm if you are going to convert your data into an interval of the data set and then like if you are trying to like uh convert those things into a sequences and if you are going to use lstm that will work for sure and as we can see it's a univariant data that we have right so in that case maybe i can try to opt for any kind of a time series based model as well even series is nothing but it's a regression model but there it's not like it will be y is equal to mx plus c so their calculation will be different so here we try to like uh find a relationship between an immediate dependent and independent variable but in case of time series so calculation happens in a different way so time t t minus one t minus two so when we combine so it's a regression model but relations will be different with respect to a time okay so as you talked about like uh lstm model right so ravi i would like to know one thing that uh like why i should use lstm so why not cnn where cnn is going to fail and where i should use lstm or maybe rnn based model um okay so uh so before like answering uh this uh let's take a step back and see like where where all these kinds of neural networks kind of fit in right so so whatever on the evolution has been happening in this space i think these they have stemmed from basically three sort of basic neural networks right so the first one is the uh obvious the perceptron one the simple one right and then we have the cnn and then we have like rn right so based on uh what i've seen so cnn is most uh suitable or useful for we can say image data kind of thing because i think cnn helps us in finding again i haven't worked with vision so uh pardon me if i find i just try something maybe you can stop sharing your screens fine so like uh let's stop sharing the screen and then we can discuss so like uh uh let's suppose i'm just uh talking about the variance of rnn which is lstm group whole whatever right now uh like uh let's suppose at the same point of a time if if i'll ask you can you please implement these things with cnns so what will be your answer and what will be your argument on that uh just a second let me let me recollect the thoughts so obviously we know that cnns work uh better in terms of vision and image tasks right but uh but the whereas the rnns or various uh what do you say the various versions and the successes of rnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnnn lstns and gru's and all they work better on the sequential data right okay but ultimately if i'll uh like i talk about the cnns right so guys i'm sending a array over there yes we can even in rnn so we are trying to do the same thing so why forget about that uh converting their images into pixels you know they are directly giving you arrays now uh yeah so let's suppose i'm talking about simple array right so even image-wise i have an array and even this data-wise i have an array right and even in both data i have to find out some relations right now tell me so why i can't use cnn or i should not use cnn in case of this kind of a data set where rna comes into a picture so how how you can differentiate this rna i'm just looking for that so i i have a bit more information regarding the rnn so i can uh because i've worked a bit more with the sequential data so i'm not really sure about the cnn but i can give some info on the rnn because rnn basically the recurrent neural network right so uh so what how it's different from the normal neural network is like it uses the it basically feeds uh the hidden state and the outcome and the output from the previous layer uh into its cells and uh this uh sort of forms a recurrent relation right so you're saying okay so like sorry for interrupt interruption so when you are saying that it tried to forward a hidden state to the next rnn cell right so what is the meaning of that so what kind of information it is trying to propagate from a previous cell to the next cell so suppose uh uh let's uh let's take the example of a text right suppose that we have a text so um so in one pass we go through the text and find some information and write so obviously uh suppose the if if i'm given with some text and i want to say summarize the text right in one pass of the text i'm not going to understand everything right so suppose in one pass maybe i understand some part of the text right and then in the second pass like it will not be able to understand everything into a single pass is it a case because i can i can say that this is the same kind of a case i will be able to get even with the cnns right so cnn will not be able to understand whole relation whole relations in a single iterations right so that's the reason so we keep on sending a data set for the multiple iteration for a multiple epoch and then we used to train it right so i think that is a very general cases with all the machine learning algorithm mostly and for neural network every new network does the same thing right so it will it will try to calculate the loss then it will try to do a backup propagations and then we try to tune the weights in a backup propagation again it will try to adjust adjustment based on the losses which i am going to receive right but my question over here is that why rnn why not cnn for this particular use cases that we were talking about okay this is definitely a tricky one okay uh so fine moving to the next one no let's let's just uh give it a few minutes so okay sure sure please please i'm just trying to have like uh let's have a little discussion about the same like because it's a fun one okay you can take help you can you can just google it if you are looking for some hint and i'm completely okay with that we can we can discuss we can discuss about your thought process no no no no googling i'll just uh i'll just discuss with you guys yeah okay so the idea which i have about cnn is like uh suppose given an image it's gonna convert it uh into say array right an array of numbers with each array representing different information about the image right and then i also have some idea like it uses some i think padding or that kind of a thing cooling padding yeah cooling padding right so that depends upon us how we are going to create the architecture that depends upon us so we can it's not material that you should use it maybe just use the convolution and then done so maybe there is no like a rule that you should use padding or pooling that's fine so uh so that hap so that thing happens and then basically with the cnn we are able to like find out which uh which parts of the image are more uh which parts of image should we focus on more right i think is i don't think so okay okay i'll i'll accept in that case because i'm not in the cnn okay move shall we move to the next question and the answer and move it's your miss it's fine tell the answer yeah difference between c and rn and right so why is c environment so these are differences right so in case of cnn it's true see whether it's a rna whether it's a simple normal neural network or whether it's a like a cnn any balance of cnn which is 16 or maybe a rest net 50 restaurant one zero one less than 30 things or maybe like a inception at google net dark net whatever right or maybe unit you can try to use so in any kind of a cnn waste or convolution based network so we are trying to take array that is fine but at the end of the day let's suppose if i passed one images and i have received some output wrong or right doesn't matter if i receive some output it will not be able to remember that what kind of a data i have passed previously it will not be able to remember at all right it will just take the images whatever wait it has learnt or it has not learned right based on that it will give me the output right and it will not be able to remember what i have done last time right but remember yeah it will not be able to remember now let's suppose me and you are trying to do some kind of a conversation so me and you are talking about something or maybe i'm trying to give some answer based on what based on the previous context just one way that we have started this conversation and based on that conversation i'm trying to give this answer because how or why because i have remembered right i have remembered what a context i have remembered i'm not i don't even remember or what you said exactly right but i remember a context in which we are talking about right and based on that i'm trying to give you some answer right or other people are trying to participate in this conversation in a chat box yeah so basically in case of rnn like you said that it will try to propagate a hidden information to a next state or a vector to a next state so it's nothing but it's a context vector or it's a state in which we were talking about so it was trying to store and will be able to transfer it or propagate it right so that's the reason so in case of this a sequential like in case of a sequential data or maybe in a time series data because time series data so always you will be able to find that it is going to repeat a itself after some time right so it will try to remember and learn the context it will try to remember and learn the context or you can say in case of regression right so what it does it will try to learn the context or pattern right one after another so what is the relationships which i am able to get so arguments are good while learning a context or propagating information from a t minus one times time to t or t plus one right whereas cnn never remembers anything and that is a that that is only like one of the uh cases but yeah so that's the reason so where we can try to distinguish between cnn or rn now coming to the data wise i can try to pass same data in a cnn i can try to pass same data in rnn it's nowhere written that inside a cnn i can only process a images it's completely fine i can try to because we are actually we should not think in terms of image or maybe a speech data or maybe in terms of sequential or non-sequential data it's just a neural network it is having a weight weights are nothing but it it's a function it is it's you can treat you should treat these things as a function weights all the weight connections that we have right so it just try to tune itself or adjust itself based on the relations now based on the loss it will try to like uh so i think i think i have made my since uh think very clearly and with respect to future extraction cnn is more powerful than rl yeah the reason behind uh is we have a convolution we have a convolution so even some of the data said where you don't have a better feature so let's suppose i have given you a data dump right and let's suppose if you don't have a better feature right you are not able to extract a better relations right let's suppose the complete feature is messed up because if you will talk about the image part right so image uh like itself try to generate a huge number of array depends upon the dimension of images but only some part will be a repeated one in all the images let's suppose if i'm talking about classification or maybe detection or segmentations or maybe tracking what are polymer statements so what cnn does so inside the cnn so we have a kernels we try to apply a kernels right learnable kernels so it will try to extract only those features which is a relevant one and that through with the help of learning so it learns which kernel or which feature i'm supposed to extract so you can try to use cnn for a feature extraction with whatever data that you have not only just with the image data set so generally like we have a very wrong perception that we should use cnn's only with uh like a image data set which is not a true you can go and you can check you will be able to find out hundreds of example on uh that one yeah i think this this particular right okay so now like uh second question i have for you so in between when you are trying to give an answer so you have you said that that for text summarization you have used a word called as summarization in between while you are trying to answer rna so we can try to use rnn for us summarization and i think you have worked in uh nlp part right as we have mentioned so uh like uh for text summarization let's suppose i have a complete book right i'm reading one book or like i have a complete news right news and let's suppose if i have given you uh this complete news and if i'll ask you okay fine do one thing so try to understand this news i'm not in a mood to like read out the complete uh newspaper right just tell me a summary of it right so which network you are going to utilize over here for summarizations and why what is the reason behind using which is your favorite there are many networks right there are many uh like an lp based network but which one you are going to use with reason okay so um so let me first break down on this whole thing right so we so the task is to like perform text summarization right yeah and for text summarization we can we can go with either rnn or the lstm or transformer based models or this any of the related uh models right okay so so out of these uh which one we shall go ahead for the text summarization right yeah so which one like uh see there are like a hundreds of models right and uh you you are working into nlp so which one you will prefer to use according to your experience this is uh only thing which i i would like to know so even last month or even this month itself on 11th of january so like google has reached one research paper uh for our trillions uh with the trillions of parameter and right i'm not able to recall an exact name of that research paper which google has read on 11th of january i was just going through it right uh and it is having like a trillions of parameters so i'm just looking for a network that you you have used or maybe you are going to use it the network that you are aware of okay so obviously we can use the rnn or lstm in this case right but the problem with the rn nlstm is like rnn suffers from the vanishing gradient problem right and lstm or gru's they have tried to sort of mitigate that in some form but obviously uh what we've seen is uh lstm are a bit of a complex and still even after uh say a limit of length they are not able to keep the context right and uh so in order to mitigate all these we have the transformers and the attention mechanism right so for summarization i think uh we need to we need a bit of a context right so like i would so i would like to stop you over here so like you said like uh rna is a complex network right or you you have mentioned that it may face a vanishing gradient issue right so can you please elaborate more about like a comparison between this uh like rnn and you talked about an attention-based model as well transformative model what i believe is like a transformer or attention-based model are much more complex right their variance they're variants of transformers and there are many variants of like a self attention or retention-based model so they are much more complex than other men so then why i should go for transformer and attention-based model because they find the uh because they are better at understanding uh underlying language right so the way they are trained the birth and all their the part is trained right so it happens it uses basically two things right the mass language modeling and like basically given a sentence it's gonna mask some words and while training it's gonna predict like which will which will be the missing word right one of this and uh the another is the next sentence prediction like whether given one sentence and given the second sentence whether the second sentence follows the first one so i think using these all all of these sort of things we get the context of the language in a much better way okay okay so what is what is your like a thought about uh how much do you know about transformers like you have mentioned transformers you're talking about bird as well right so what is birth why what why what is better than uh this one and before before even like uh talking about a birth so can you please try to like uh highlight what is attention-based model or let's suppose self-attention yeah sure so so i think we should start from the beginning like how like what is the need for all of this right like starting from the let's say the most most basic uh ones right so say we have bag of words right so so there's this bag of words and then why do we need to go from bag of words to like something like a word to work right so i think after bag of words dfidf i think what to work was uh one of the big things so so back of words in tf idf uh they obviously they don't take any relationship uh in between the words into account they don't uh like like suppose if we are representing text in form of bag of words or tf idf and suppose we have two words like uh tiger and line so in no way it's gonna capture the relationship between these two words because say we plot these two words on a graph so they should be close closer to each other right because they are sort of related so in order to capture the context or relationship between the language or the words better i think then we went with the what to wet model which which had like one neural network to capture the context right but then again it wasn't able to do that in a full-fledged manner because it's just one shallow neural network and it failed in basically it failed in cases like suppose apple is my favorite company and the second case would be i eat apple every day so in both the cases apple means different but uh an algorithm like what to work it's not gonna differentiate right so after uh so after this we have uh uh so even between i think what do we can but they were like elmo and ulm and these kinds of embeddings i'm not really sure about the other ones so then came the transformer and the self-retention ones right so using so if we talk about transformers what i think it's it's some kind of an encoder decoder architecture right so uh based on that what happens is is like um uh what do you say another pro or uh advantage it has another advantage it has from the previous ones is that we can pass the data uh in parallely so that uh takes uh so that takes advantage of the parallel compute we have nowadays right so coming to the transformers uh architecture so say we have uh transformers and in that we have uh would you say the encoder and decoder part right so we had encoder decoder right and inside encoded decoder so we can try to use maybe like any balance of rnn so maybe like a stacked fgm or maybe like a bi-directional or a gru or maybe like a peephole anything i can try to use right right we had encoded equator right so why uh we are moving towards the self attention and then transformer okay so uh in self attention what happens is uh i think uh using self attention we are able to sort of attend to various parts of the sentences or say pay attention to various parts of the sentences in a much better way than some examples so let's suppose uh i'm a completely like a non-technical person i don't understand these things and uh you have to explain me so how you are going to explain me okay so okay for a non-technical person yeah i'm completely like unknown to uh this transformer you have to explain me problem and then uh you know layman like uh solutions by using this attention you can also take out the architecture and probably share your screen ah if you want here so i don't have an issue no no i don't let's just discuss it like this it's okay it's okay to have architecture in front of you you know so you'll be able to explain more better it's your it's your choice okay so um we were at the transformers right i kind of lost track so okay so we are at the transformers and uh first of all what is transformers right so in layman terms it's an encoder decoder architecture right and uh the new ones use the self attention in that okay and in encoder and decoder we can obviously use rnns or lstms or grdu's any of those things right but why we use self attention uh is because using self-attention we are able to attend to various parts of the sentences say we have a big paragraph or a big text and using the self-attention mechanism we are able to uh find out or basically attend to the text in a better way than uh rnns or lstms right because uh it's it uh giving like going more into a bit of depth like we have uh we calculate like various types of embeddings like there is positional embedding and then there's a segment embedding and has been word so like they have introduced this concept called as personal embeddings right so that it will be able to relationship between the positions and then it will try to insert the data inside along with your word embedding so word plus positional binding combine it and then send the data into a network uh which is not so like a transformer will come into picture so this will be the input that is fine so i'm just trying to understand like uh your understanding of uh like uh why is what is self-attention what is the need of self-attention right we already had a encoder decoder architecture right so why they have modified this architecture again if i'll talk about like uh attention or maybe transformers so maybe in some of the transformers they are just using encoder layer inside that so again we have a multiple layers right and as in some of the architecture they are using encoded equator both somewhere they are using decoder paths right so just in in one single line if you have to explain why self attention with example then then what will be your answer okay so in one sentence what i'd say is given any a text of say any length we can find the context using this uh using self attention in in the in in the best way example in any example still i'm not able to understand right so just give me some example to uh like so that i will i will be able to correlate i'm a person who is very weak in like a coordinating things uh so just just give me some example okay so let me share my screen so i can yeah you can do that share some image right yeah okay is it visible uh yeah okay so i think this is the picture that comes to my mind when i'm trying to like correlate with attention so basically using attention what we can do is basically uh say suppose we have a corpus this is the corpus right this is these are my sentences right now the model doesn't know the language like humans do right but using the self-attention what it is able to do right it it's able to figure out like what it is referring to okay so using attention we can adjust the weights of all these words and the and this is do and this is like we can say multi-head attention and that combination right so we can have various attention depends upon my architecture so it will be my choice looking into my images you can't say that it's a multi-year attention right right so uh so with the tension uh what we can do is like see now the word it before uh before doing our thing with the attention and that we weren't aware like what it is referring to right but after attention and after the weights have been calculated it is aware like so the basically the model is aware like what it is referring to in this case okay see okay let me like uh tell you uh in a best possible way see what happens over here is attention right so as a word if you try to go after the meanings of this word attention right so it's a meaning is very simple so meaning over here is i have to generate something right or maybe i have to convert i have to do a language translation this is the task that is in front of me or maybe i have to do a text summarization it can be abstractive it can be extracted like a summarization so whenever i'm trying to do uh summarizations or whenever i'm trying to like convert one language to other languages whenever i'm trying to do a question answering right this is like an llp task so whenever i'm trying to do a question answering or some kind of a sentence or maybe what or maybe character generation so generation of that particular sentence right depends upon so it's whenever i do a language translation so it's not like it will be one to one mapping of a language translation right one word which i'm trying to generate let's suppose hindi to english or hindi english to french right or any other languages if i'm trying to generate so whatever word that i'm trying to generate it could possible that this word generations depends upon many other things because machine never understands grammar right machine never understands your context right so attention based model is designed for a purpose so that it will be able to focus on all the word that we have or all the sentences that we have and it will try to learn to generate one particular word which are all word on which i am supposed to focus now coming to this particular picture that you have opened up right from google so here as you can see it right now there is a word called as it now it may represent or let's suppose if it is it is trying to generate right so it so generate this particular it it may depends upon all the different different words which is available into my sentences the animal didn't cross the street because and all so on right now here you will be able to see a different different color contrast right so you are able to see a different color contrast it simply means that it's just a representation diagram now if you will go ahead and if you try to open up research paper you will be able to find out the same thing so it simply represents over here that what is a weightage or what is a contribution of this word to generate a next word so it is just showing you our attention means if i have to generate it right so this generation will depends upon which and all word if if i'm talking about one particular scenario where i have to generate some word right so attention simply means that that okay i have to generate something i don't know on which one all i have to focus i learned it so what they have done so they have placed our attention-based network or neural network in between putting up some kind of uh rules in between for weight initializations of those things if you try to study that like a first self-attention research paper you will be able to find out over there very clearly right attention based functions you will be able to get you will be able to see what was the conditions they have given on await and all sort of things so a simple meaning of attention is i have to do one single task but to do particular task i have to focus on multiple things and all the things a priority of all these things are not same so what i will do is so i will try to learn to complete this task what in all priority i can try to set again in a layman way if i have to explain right so let's suppose if i'll ask you to solve one particular problem right one particular problem so here i have given you one task in the very beginning saying that okay fine so just try to load this file and then try to perform this task based on the data type filter it out right now so the final task was filter out right that particular record based on the data types but to perform all of those things you have paid attention on many things like loading a data set right checking what is a data type and then like probably writing a multiple like a lineup for a different different kind of a queries basically breaking down the problems related to all of these tasks that you have right but again priority of every task is not same right so now you know how to import up and down how to load a file so priority is different right and it's not like you have involved every concept which is available inside of pandas there are thousands of concept there are thousands of function which is available but you have paid attention to only those functions or only those approach which is related to your task to get your output correct right right so you have used pandas but it's not like you have used the complete pandas package you have used all the apis of pandas i haven't used like entire library i only focused on what was required you have paid attention on couple of things by which you are able to achieve your task right so same thing which people have introduced over here that if i have to generate a new sentence right so why can't we build a particular model why why why can't we train or why can't we place a particular model which will try to focus which will learn it's a neural network so fortunately we'll learn based on the losses right so it will learn that okay where i have to focus and based on that probably it will be able to give me a best possible outcome right so this is uh in a layman way this is how uh maybe i'll try to explain right with respect to your task yeah you can like unshare your screen that's fine okay so chris you have uh something yeah now i just have one simple question okay if you're given an nlp task so what will you be your approach to solve it like will you go with directly with the state of art algorithm but why but why like your state of art algorithm gives you very good accuracy right with respect to anything see so i think it's always first better to start with the basic approach and then build up and then sort of add the complexity because what if in my case i could have used some simple tfidf stuff and then and then that would have given me uh sufficient results right why would i waste probably tomorrow you have more amount of data that will be coming up probably tomorrow you may have some new data that will be coming up in your project and probably by using a state of art algorithm you know you'll be able to get a good solution so why not so even even in that case i i think uh what my approach is in is i think we should start building the solution from the bottom up like the least uh possible complexity and then keep adding the complexity and like like if i'm not happy with the results i can increase the complexity uh right and also my learning will be better in the first case because since i'm doing from the since i'm moving up from the basic approaches so why do you think these companies are coming up with such amazing state of art algorithms then yeah still if you're using this old uh technique or libraries right really i'll say it's just a kind of a dictionary right it's just a library built on top of most of the dictionary maybe in childhood we were using a oxford dictionary and now it's available over here with some additional functions right and some a pre-trained network as well if i talk about space agency or maybe a stanford nlp or maybe an lte right and all sort of things so why why people are investing a lot by like open ai uh has released this uh like a gpd3 and on january 11 2021 this year itself i think 20 days back right so google has released just to compete with a gpd3 so google has released something called as switch transformer this is the research paper search yeah so why they have released this model and why they are releasing this research paper if these things are not relevant because what i believe is like it takes a huge amount of investment right and if people like you are going to suggest me that no i should not use this one right i should go ahead with the conventional approach now i'm not saying not to i'm not saying not to use this tomorrow you may be going tomorrow you may be going in a company right uh probably you'll be working in that and suddenly what will happen is that you will get a task then how will you convince your manager which approach to go your manager will say okay let's go with the state of art algorithm google has come up with trillion of parameters let's not say that let's go with the bottom move why why will the manager wait for that much time he'll say that i want the best accuracy yeah so if if he says like he wants the best accuracy and he doesn't have time then obviously it makes sense to go with the best rates now what will be the problem if you directly use the state-of-the-art algorithms probably yeah so the second thing is so algorithm that we are saying as a state of the art algorithm right and which people have uh given like for which research people are available and even bench markings are available on a certain data set right not on your data set so it could possible that uh it will not give you a same accuracy that those people are able to receive it and most of the time so it's not easy right like your like library or maybe your conventional uh based things so it's not as easy to implement right if you try to compare these things with your conversational lp uh so what will be your argument in that case yeah so this is a tricky one and again productionizing this this this smaller expansion right but elmo gpt-1 gpt to write a t5 electra right and then this uh gp3 and then this new model right switch transformers so this this looks fancy right so we all uh are talking about it but again so if you will try to productionize it so it will be having its own uh like uh hurdles it will be having its own complexity so what will be your suggestions if you just try to load a trillion parameter model just you know what will happen how much response time probably it will take right which problems you'll be facing in order to do the deployment how much dependency you'll basically be having right so usually we don't go in that specifically unless and until your company is developing something that needs to be provided as a service to others and again so everybody is a huge issue chris because like i said right i told you i was just one of the use cases that we are building so uh like uh scalability is uh again a huge issue so maybe for one instance one model i can try to host it but the thing is like let's suppose if i have to serve this things to 5000 customers in a second so i neuron is developing one product where they're using megatron okay and i think he's facing issues in that so we ask you this question uh like we are trying to use a state-of-the-art model we have experimented with almost like every model every possible model that you can heard of or that has been released till december 2020 that that was released right so uh for whatever paperwork we have converted it and uh yeah it is it is having its own uh now i can say like uh complexity and huddle so what will be your take on this one right okay so obviously um productionizing these uh heavy sort of models is is gonna cost a lot of compute right and uh even if i have these kind of a heavy models and if i have plenty of requests coming in and then in real time the latency is going to be high which is not a desirable position to be in right so another solution yeah let's see accuracy part is this accuracy part is again one of the factor because whatever accuracy they have claimed it doesn't mean that you will be able to get a same accuracy with your data set right they have already mentioned a squad data set fine so let's suppose if i have done bench parking or squad it doesn't that is fine that was a squad right hours will be different hours will be a different uh parameters and then like different different warnings and different different vectors that is fine so have you have you uh have you heard of like a spark nlp have you compared spark nlp with uh any other existing nlp framework in a market so uh i want to say like uh two things regarding this so uh one thing is uh there's uh one technique called i think distillation model distillation or model pruning i think so but in model pruning so there is a huge uh there is a very high chances that maybe you will end up losing some of the relations so again accuracy will go down okay so another model distillation kind of thing that i think that means is the sort of compressing your model uh so that the size becomes less than half but it's gonna so i think that's distilled but i'm not i'm not really sure if they productionize it or but but they say like it's less than half the size of the model while still retaining around 95 percent accuracy so that's one thing so another thing which you mentioned is park nlp right yeah i've i've seen this library it's by i think john snow labs right so so it's it's not like uh i won't say i have used it a lot but uh i've seen it and they have implemented various named entity recognition and sentiment analysis and all those uh models so yeah i've uh i've heard about it but you use it so my solution to you will be instead of using this conventional uh like a libraries like stanford nlp or text blog or spc or maybe so try to use a spark nlp uh hopefully you will be able to get uh very less latency over here and even accuracy wise so wherever like uh this model this library is used to involve a prediction or model based like a prediction an outcome so it will give you a better one but not better than a state of the art yeah but it's fast it's fast because of its like optimizer which has been implemented in a backend of a spark engine so yeah because of that so yeah i think i think this is it from my side chris so like uh i'm done so you have anything yeah no so i think uh that's it we have extended more than one hour probably this is the first interview tomorrow we still have many things planned but i hope any feedback that you want to give sudan show to ravi uh you can go ahead and give uh yeah so my opinion uh for you will be like uh try to get into like a architecture in depth like because whenever uh because in uh data science right so when we talk about like starting from the very beginning so when we talk about uh like a python itself right so maybe you are using uh this pandas library so if i'll start from there itself now if i'll give you a problem statement that let's suppose to read out this fifteen thousand data your pandas library is using maybe 100 millisecond right now i have to bring down these things to maybe like uh two millisecond or three milliseconds to read up this much of record right so maybe like uh you'll you'll have to like use some other libraries maybe task maybe a spark based solutions you can try to use or even in machine learning deeper multi threading maybe which one multi thread i think by multi threading we can ah no so by multi-threading so you will be able to make these things maybe 2x faster 3x faster right there will be a limitation with the processes with which you are going to define the threadings threads but here it's taking let's suppose 100 millisecond and if i'll ask you bring it down to two milliseconds yeah right that's that's a lot that will not help you out right and you'll have to think of some other framework yeah so uh again that was one of the questions one of the very interesting questions which was asked in a recent uh interview uh because we all know that uh processing our data is a huge hurdle for every one of us right but um one thing which i've heard regarding dusk is that if you do some group buys and if you do some operations it consumes more memory i think uh yeah it is going to consume a more memory because it try to store a planning for those things right right so again this park does the same thing but in a spot so there is something called as a catalina optimizer which is implemented in a backend and because of that it optimizes those complete queries but so yeah so like uh so there is many things right so there is many things like in machine learning and deep learning computer vision nlp right and we as a solution provider right so we always are developer or maybe like a people who are trying to design the system we should know that that which one is good which one is bad which one is best right and what i believe is in this comparison comes from an architecture plus implementation so if we know the architecture that is not sufficient unless and until we have not seen it because many people used to say that okay fine so uh why can't we use bot or maybe this hugging face api just rightly call it ultimately it is calling those state of the art model itself right right so use bot or maybe t5 or maybe use a distal bird like you said right or maybe for embeddings use elmo or something else yeah but thing is we should know the restrictions in terms of implementation in terms of latency in terms of our resource utilizations plus we should know architecture because architecture always helps us out to think through the entire like things right and this is uh where i believe you should you should look into this area right and uh yeah what i have seen is like uh your coding part is pretty good uh so at least you are able to think the way you were like uh writing it right although it was a very small uh like a problem statement that we have given to you right but the way you were thinking it and thinking uh through these things and then we have implemented anagram or whether it's a simple extraction so i think uh you are good with that i if you can focus on this complete architecture plus comparison plus implementation part so things will be like a better in that way yeah so that will be my like uh opinion you can see thank you sulanjo for the feedback and again yes ravi from my side i think your basics is strong uh with respect to coding and all uh the thing is that uh you know with but now since you will be getting experience now as you go on working with data science probably now you need to think about why how what at every instance of time you know because uh after some years of experience you have to take decisions so at that time you need to understand like why you're taking that specific decision but i hope you liked this interview there were a lot of complaints and previous interviews we used to not say the answer but i think this time right both a lot of answers were there more than questions right with along with the explanation so i hope people also like to who are our audience they also had fun even when i used to watch your previous interview experiences right and even even in those you you used to ask some really tricky questions but the answers were not uh kind of available so so that that was the because at the time as soon as you clear go after that enter you know you will be thinking okay why why why that will at least irritate your mind away you got all the answers probably after this you'll be relieved and you'll be keeping right so to just keep that thing going on right we should not say that answer but many people wanted the answer so we thought of this interview so like a uh like we are we are not supposed to uh in a real time interview we are not supposed to say the answer but yeah it's uh for learning uh for everyone right so uh what we have seriously okay fine so let's let's give an answer for uh each and every questions that we have asked so that our list may be hint right so that they will be able to like uh go in a right direction they will be able to uh explore yeah and what i liked about you ravi is like uh at least you uh like you are able to think in that direction but there is a lag so if you can work on that so probably in next six months or in a year so you will be uh able to like answer all those things logically and in a correct manner this is this is what i believe yeah yeah so i had also based on previous interview experience when you guys asked about which is your favorite machine learning algorithm so i prepared that one specifically but you didn't even ask that sometimes whatever you prepare that does not come in the interview right yeah the right example for you so everyone says random forest but i prepared for linear regression this time but whatever okay again sometime we'll have a discussion regarding that again if we talk about linear regression that will be a whole one-hour discussion again separately you know so uh anyhow i hope everybody liked this interview i hope you liked it ravi uh you had fun uh yeah it was knowledgeable you learned a lot you know and i hope audience also liked it so yes we'll be continuing this series for one month at least uh and every day you'll be seeing us especially me and sudanshu so thank you ravi i hope you take up this feedback and keep on working on it okay amazing yeah session and they have uh gone through it yeah thanks very thanks a lot yeah thank you thank you everyone this was it from our side for me and sudan show i hope you liked it along with ravi hit like for ravi i have given the linkedin idea of laravey in the description of this particular video you can actually communicate to him you can chat with him you can ask him that how do you know this much things okay that will be helpful for you all uh and yes we'll see all in the next videos have a great day thank you all thank you so