Generative AI RoadMap: From Beginner to Master in 2024 #generativeai #artificialintelligence

[Music] [Applause] hey there everyone welcome back to my YouTube channel so guys the wait is over finally and firstly thank you all for the great success of my mlops video it's more than 2.5 plus views and more than 100 likes I did not expect that it is going to reach that high so guys now I don't want to stop so I have created a complete generative AI road map and I'm expecting that even this particular video is going to definitely reach more than 1K so the target for this particular video is more than 1K views I hope you all are going to definitely like this particular video and my road map which I have personally set and created for you all and with the resources I'll be providing you all everything please make sure you subscribe my channel and share this video to your friends who want to learn generative Ai and llms it will be definitely helpful for them so guys now let's not waste our time let's dive into my computer screen and discuss more about this particular road map if you have not subscribed the channel do subscribe the channel and on the Bell notification and you all can definitely connect with me on LinkedIn and my LinkedIn ID is below this particular video you can check out the description so now let's dive into my computer screen so guys this is the road map which I have created for you all where everything is explained clearly I'll be explaining in detail make sure you all have a patience and take your coffee or tea sit back and just watch this video completely you'll definitely love it okay and I'll be providing you all the resources also for each and every topic don't worry so first let us see about the basic maths so if you want to start generate AI I'm assuming that you should know machine learning deep learning and NLP and I'm assuming that if you want to start machine learning you should know the basic maths which is linear algebra calculus and Pro probability and statistics make sure you'll have a basic idea about this basic idea in sense I'm not telling you to sit and you'll be doing some calculations and all mathematically if you know how it works and if you know the logic behind those topics it is enough because the machine learning algorithms will do everything for you all okay so guys I'm assuming you all know this particular topics once you cover this particular topic then you can move further and to learn learn this particular topic I have created again a GitHub repo how I have uh in my mlops video right there was a geub repo which explains the structure the same way you have this particular GitHub repo I have taken it from a person I have cloned it from Melbourne llm course main so I have taken it from him and I have modified some things here because this is a nice road map which he has taken but there are some things which are not updated which I have updated in this particular particular road map video so you can see here even this road map image also you can check out but here I have explained completely clearly and the structure is clear if you you can follow this one because there are some skills which are missing in that particular road map I have added here okay so for basic maths you have this something called as three blue and one brown uh video okay this is a YouTube channel and they have explained it very clearly and even KH Academy you can check out all this videos the resources are very good and you can watch those video you will understand the basic maths very well okay so guys after this I'll be opening this particular uh repo again and again and I'll be explaining everything clearly don't worry okay so after reading this basic math then your first task is over after this you need to know machine learning so in machine learning you should know python Basics data pre-processing and machine learning library so guys if you are regularly following my channel I know I guess you all know which video you have to check out for machine learning there is already a machine learning road map video and I'll be updating that video uh Link in this particular video description you can check out that particular road map where I have given a proper breakthrough how you should start machine learning and which algorithms you should learn and which are supervised unsupervised everything is explained clearly you can check out my machine learning road map that road map link is in this video description okay and same comes for your even deep learning so in deep learning basic DL which is nothing but you should know about you can see I have mentioned here already so you should know about the fundamentals like bias variance weights activation function tan Theta tan Theta Ru function and then sigmoid everything all these things you should know and the different training and optimization techniques like mean square error cross enthropy gradient descent sastic gradient descent algorithms RMS prop and Adam all these things you should know then when it comes to overfitting for overfitting what you're going to do you should use L1 and L2 regularization early stopping methods and uh data augmentation so all these topics you can check out again there are some videos related to it you have free code cam fast Ai and even three blue one Brown video also this video resources are very good you can check out them to understand them in detail and you should mainly know what are perceptrons and multi-layer perceptrons okay so about them and once you know de deep learning you should definitely know the Frameworks we are going to use for deep learning in machine learning we are directly going to import the algorithms and train based on that but here you should know no tensor flow or P torch okay these two are the Frameworks which we'll be using in neural network so you should know them and you should be able to create the architectures using tensor flow or py toor okay so you can use any of them in neural networks mainly make sure you all know a Ann properly again there is a deep learning road map where I have given the complete breakthrough and explain detail about deep learning so it was released I guess last month in a sense it was released on January 2024 only you can check out that particular video that video link will be in this video description okay it's a latest road map after that NLP so you should know NLP after deep learning which is text pre-processing you should know tokenization feature extraction word embeddings and RNN also you should know some of the NLP architectures like Transformers attention is all you need encoder decoder so all those things also have been mentioned here I guess yeah TF IDF engrams byrams all these things you should learn okay so for that also there are some videos out here and J Almer videos are there I would mainly recommend you all to check out J Almer videos and his website also it is there if you search for J Almer you will get his website he has explained everything very clearly and lstm bstm Gru about all these topics I have already created a complete NLP road map also for whatever you're going to see here already there are many road maps which I have created okay so for NLP also there is a road map you can check out that road map I would recommend you all to check out that road map okay so guys I me that was released before this video only last week only I have released that particular video you can check out that video okay and once you complete NLP now you're going to enter into the battle so sit back it's not a hurry at all so make sure if your coffee is done you can get one more coffee or juice I'm just kidding okay so once you done with NLP your Basics are clear and strong okay now we are going to go with lm's architecture so we are before we dive into this architecture I have an announcement for you all so I used to get many requests from many of them like they are messaging me in LinkedIn and they asking me many doubts related and the resources how they have to start learning and what issues if they get how can I solve them so there were many issues also some of them were having issues which I have personally took them on LinkedIn call and I have solved them but always I'm not free to connect with them right so guys I thought why don't we build a community so I have created a Discord channel for you all this is our Discord server or Channel community where you all can join here already many of them are joining in this particular server you can see here there are many of them who are joining to my channel and there are some welcome messages and rules as soon as you join you get it you can read this and you should definitely follow this rules out if you don't want to get banned I'm not going to ban you all or else I'm not going to remove from my community there is a bot called you can see this is the bot okay server supervisor this is the bot which is going to monit or everything and if there is any spamming or else any illegal activities in a sense like any nonsense activities then it is definitely going to remove you from my server okay so guys make sure you follow that rules and there is announcement session I'll be announcing the latest uh uh videos whatever it is there latest updates related to my channel and the community there are many llm updates so which I will will be going forward sharing here and even job referrals so you all can join here and if you have any openings in your company you can definitely provide the reference to the people in this community and even if there there are some people who want the jobs even I'll be posting the job posts very soon and there is something called as referral session if you have any openings in your company you can post the message here those who are in need of job they'll definitely contact you from the community okay you can even make the personal connections around so even you will get the resources if you ask for the resources we be updating the resources here okay and when it comes to the personal chattings also you can do and even the group chats you can talk with the people around and very soon even I'm planning to post a normal voice chats around in Discord every week we'll try to have a general normal conversation with my community people and my subscribers based on it I'll get to know more about what are the latest trends and I can learn more about it so guys make sure you all join my community and now let let's move further with our video I have already created a detailed video about my Discord Channel you can check out that video also it's in the channel so llm architecture Transformers so you should know Transformers which is nothing but you can even call attention is all you need there is something called paper attention is all you need which was introduced by Google so Transformer is based out of it and there is a video okay Transformers NLP explained if you just search for this from Google there was a video yeah this is the video which was introduced two years back but this particular video without any coding or anything she explained everything very clearly in 9 minutes you will understand about Transformers GPT bir and T5 what is behind working procedure you will easily understand you'll get the basic idea about Transformers I would recommend you all to see this video and even the person code Emporium you can see right his Transformers uh explained neural network that particular video is very awesome you can check out his channel also okay so guys after uh llm architecture then you should learn instruction data sets which is nothing but Al Alpha it is Alpha yes I guess Alpha and eval instrict so about this you can see there is in my channel okay so you learn about those llms okay clearly and attention mechanism and text generation techniques Alpha like data set which is nothing but this is used to generate the synthetic data from scratch with the help of open AI GPT so you can use this if you directly open this you will get the overview about it and how you should use you will get detail overview about it it is going to use text DAV 003 and Lama 7 B okay to generate the models synthetic data okay I guess now text instead of text generation 003 this is deprecated right now this model is deprecated instead of this there is something called as jpd 3.5 turbo instruct okay you can search for it if you open the open AIS model you will get it so make sure you can see about that and there is something called as another Advanced Techniques like how you can improve your existing data set for that you can use Evol instruct and to generate high quality synthetic data you can use pi1 or Ora after that filtering data for filtering data you can use the traditional method like Rejects and removing near duplicates and other techniques are there you can check out them rejects you can directly import re and you can do it in Python it is in I mean there is a library re you can import and you can do it and for promt templates you can check out chat ML and Al Al I guess I guess I'm pronouncing properly this one and this is that one only which we have checked first only right the same one okay so you can check out about this particular model clearly after this pre-train models so llama 2 gp4 B Gemini GMA Gemma was introduced on Feb 21st okay so I'm just adding that model also because it is latest mixl 7B and Falcon mix 7B Gemma and Falcon are open source models okay so you can use them you don't need to pay them they are completely open source you can access that about Gemma I have already created a detailed video if you have not watched out you can check out I have even done the real time implementation on that that particular video will be in this video description okay so about them if you want the resources you can check out here Lama 2 which is trained on two trillion tokens and there is something called as Megatron GPT neon and scaling loss HPC so these things are very clear okay data pipelines things are there right so you can check out about them very detailly if you just open this here you will get it and remember one thing Guys these all are a pre-trained model so which is nothing but you don't need to write any code from scratch you will be learning everything like you will be just importing them and you'll be just fine-tuning them okay based on your fune data set it will this models will work okay at the back end so if you know how to use one model that is gp4 or BT or gemini or even Lama 2 if you know how to use them you can automatically use all other models the workflow is same only the structure how you are going to use that will be different that's all okay the code you are going to write right which will be that is a bit different that's all apart from that all these models are same if you know two or three models then you can automatically understand them if you have the basic idea or overview about this models it is enough okay after this you have supervise fine tuning I guess you all are not feeling sleepy guys make sure you all see the video properly and make a note down and if you don't have the interest to make a note down at least keep some snacks aside so that that you will feel a bit interested while listening these topics okay you don't sleep okay I hope if you are not like the video please make sure you like the video and share this video with your friends and even post about this particular video on LinkedIn it will be definitely helpful for the community and even I'll get the subscribers because of your one post so thank you all now let's continue with supervised fine tuning okay supervised fine tuning so in supervised fine tuning you have Laura Q Lura xlot and deep speed so guys if you check out here so for supervised fine tuning which is nothing but pre-trained models are only trained on a next token prediction so which is why they are not helpful assistants and they cannot like to say frankly they can only predict the next token okay they are only trained for that okay so for that you can you you have the models like Laura a parameter efficient technique okay uh but based on the low rank adapter instead of training all the parameters which are only used by these adapters okay and QA which is an updated version of Laura you can say and it is also uh same P based Laura technique which is nothing but pre efficient technique all these are supervised techniques guys you can check out them clearly deep speed which is uh used for pre-training and fire tuning of llms for multi- gpus and multi- nodes okay so you can check out about them in detail you can just directly open the link here and you can check about them you can read about them okay so these all are different fine-tuning techniques that's all Transformer with deep speed you can open this and you can learn the Transformers blog this blog will be very clear whatever hugging pH is going to uh like release right guys that will be very clear I have told that in my previous videos also now also I'm telling you the same they are going to explain the things very clear you can check out hugging phas uh documentations first they have in that and if you're not able to understand that Medium blogs or you have your gpts which are nothing but your Co Learners or else you can say co-pilot who are going to explain you everything clearly so now let us go further with RH rlf which is nothing but reinforcement learning uh for human feedback okay very long name right okay after supervis fine tuning rlf is a step used to align the llms like human expectations the idea is to learn preferences from humans or AI feedback so it is nothing but this means that based on the human feedback the model is going to improve what does reinforcement learning suggest so without any further data it is going to like based on available data whatever feedbacks we have going to give and whatever we are going to train it is going to learn it is going to reinforce itself the same way here based on human feedback the model is going to improve a lot okay so you can say even your Gemini and gp4 these all are reinforcement learning human feedback model so if you just search in your gemini or this one they'll give you like was that answer solv or not it will ask you to like select right if it is solved you have to like or dislike so that is what or else previously it used to give us rating now they have stopped to give rating so in this you have preference data sets uh proximal policy optimization direct preference optimization okay these data set can typically contain several answers with some kind of ranking okay we are going to provide ranking which make them more difficult to produce the instruction data sets Okay and uh proximal policy optimization this algorithm reveres a reward model to predict whether a given text is ranked by humans or not okay there is sft model for this with a plenty based and uh KL deligence you can just check out about this open this in new tab everything is available so you can just read them clearly if you just click on this download PDF it is not going to download it will show you the paper directly okay you can read this i showing you all all this it doesn't mean that you should learn everything right now itself if you try to learn this completely it is approximately going to take more than years so I'm just going to tell you all just understand the structure and start with the implementation so slowly whichever if you take any data set and if you're are going to work on that whatever is required for that data set work with that while you are going to do that project specifically you're going to learn parall each and everything so make sure you do Project based learning instead of learning all this theoretically okay for everything make sure you should have the access to open API you should have the API Keys hugging face keys so many of them don't know how to get hugging face API key right just search for hugging face API so it will just give you interface serverless serverless hosting instead of this there is a API key with hugging face itself so yeah and organization profile settings in settings it should be I guess access token so this is your hugging face API key okay I'm not going to show it you can use this access token and in building you can see like if you are using anything billing you have to update that you I have not updated that so you will not see any billing here so you have something your authentication two Factor authentication just check on your profile you'll get the details more okay so you will get hugging face API key using which you can even uh work with this hugging face API keys after this the main thing which is llm comes in llm API which is nothing but open API you have Google's now right now Gemini API is there an and core these all are the different apis which you can get and which you can use the llms at the back end okay using them kohar kohar is having a API using which you can integrate to your website so how it works right let me just give you a basic overview for example you're asking a question to like you have a website created okay a bot okay so now whatever questions you type what it will do first it will go to the back end and it will check the it will send it to the bot okay the back end bot or model which ever it is trained it is going to send there the model is fine tuned with the data set which you going to give there it will check the answer okay then it will hit it to open API so for that model if you want to run you require open API right so it will back end it will trigger the open API request and that open API will check the find T model which we have created and from that model it will pick out the result and it will prompt us the output okay this is how the normal llm apis work okay for more detail information you can check out about llm apis in YouTube you'll get many videos you have some open source llms also like like API is also like gugging face spaces llm Studio llama CP and ama ama is also open source okay this is right now latest trending you can say many of them want to learn about you can check out hugging face spacers I have already told you all I guess in aing face spaces there are many models you can directly open them and use in NLP road map I have explained it if you have not checked that you check out there I have explained clearly and prompt engineering make sure you learn prompt engineering for prompt engineering I have a direct recommendation check out deep learning.ai there is a website okay in which Andro and G is a famous Mentor you can say many of them might have known about him and some of them don't even know he is a mentor and he even teach the people in Stanford also is that great so you can check out Andro and's lecture related to prompt engineering he has explained in that very detailly okay structuring outputs lmql and outlines and guidance okay so you can see here everything is clear in this so quantization and evaluation how you are going to evaluate the model about that here it is there human evaluation or task specific benchmarks this I have not mentioned there but it is in GitHub you can check out okay Quant ization is also same now when it comes to here running llms okay yeah these are going to these are nothing but they are just going to provide you the structure output like PDF or how you want based on that it is going to provide okay for that we use all these different techniques you can check out them you can see a structure output templates like in Json format and there are some libraries like those are libraries which you using which you can generate a specific structure of your requirements okay building Vector data base which is nothing but embedding database embeddings are vectors are nothing but they are embeddings right which you are going to create in NLP many of them get confused what is Vector what is embeddings which is different or not vectors and embeddings both are one okay traditional document loaders are the option so you can see here we have written detailly like the PDFs Json and markdown files and HTML these all are the traditional databases and splitting documents like uh you can split the documents into chunks in Vector databases different different chunks you can make so for example you have a document which consists of more than 1 million words you can make thousand thousand different different chunks so because of which even the model will be optimized our llm models even the training will be very good and clear okay you have something called as splitting documents that is done okay embedding models so there are different embedding models are there right you can use them like converting the text into Vector representation for that and there are some Vector databases also so for that you can see here there is something called as chroma DB you can use chroma DB pine cone so pine cone you have the trial version you can check out even pine cone also and Anoni fasis malas so pine cone I have used it was very good and it was easy to use you can check out that okay so after this here comes the ultimate uh like ultimate topic which is rag retriever augmented generation so there are two categories in rag Basics and advance so in Basics which is AUST sorry or Chesters or Chesters which are nothing but Lang chain llama index and fast rag these are the Frameworks and using which these are like the latest trending rag Frameworks rag is a you can say rag is right now most famous topic everyone wants to learn rag even I want to learn more about rag so you can say you can learn about rag clearly so there is Lang chain llama index these are Frameworks using which you can learn for L chain I would recommend you all to go for uh you can just search for Lang chain tutorial if you just search for this you you will get this guy at First Data in the IND Indy his playlist is very clear you can understand Lang chain very soon you can check out his videos okay for Lang chain even llama index is explaining you can check out past track if you Google you'll get it if you understand one framework it is enough they are just framework implementation will be same okay and everything retrievers you have some retriever chains like multiquery Retriever and hide hyd you can see here okay so you can learn about those retrievers how the retrievers work and everything memory so memory are nothing but the llm models how which are nothing but this gpts and everything we are going to train them in a way that they can remember right so that is called memory based utilization you can learn about them using Rag and evaluation so there is something called as ragers and deep eval so rag rag evaluation techniques also are there and deep evaluation method is there you can check out about them here so these are some simplified tools which you can use to evaluate the models okay these are some basic rag applications which you should know after this you have something called as advanced rag which is query Constructors agents and tools and post pre-processing about all these topics you can explore in detail when you learn about rag you will get everything detail so for rag I guess there are some YouTube videos related to rag because it's the trending topic right now and everyone might have explained it clearly r you can check out krishna's YouTube channel he has explained it well you can check out it or else even Tech with Tim this guy also has explained very clearly or else yeah engineer this one you can check out because he's Jerry Lou I know him is a CEO I guess for llama index he will explain it very clearly you can check out his video also his YouTube channel AI engineer is very good enough you can check out Jer Leu videos llm Ops so once you are familiar with this make sure you learn about LL Ms so those who all have watched my mlops video by now you have got what is LL Ops those who don't know let me just tell you those who are hearing this word for first time llm Ops is nothing but deploying the llm model into the production you can say so that is called llm you can converting them into a website you can easily okay so that is called llm ops using Cloud storages and techniques so this uh you have to learn this it's not mandatory right now but if you are completely interested into going to operations then you can learn llm Ops or else if you know all this topics it's enough see it is completely Project based learning when you are going to implement you'll be slowly learning one by one don't learn everything one after another take one topic from each from each topic take one particular topic and learn it like for example in this in this only take first this traditional document loader learn about that then splitter document loader so use learn one one technique one by one okay don't learn everything at a time you will definitely get confused so for local deployment there is something called LM studio is there supports for local deployment ooga and cal. CPP so these all are there for local deployment in since you'll be installing them there might be a library or else even there are tools you can install them and you can directly update and run and immediately it will prompt you out in a local deployment so for demo deployments you can use gradio streamlit and even hugging face spaces so they provide hugging face spaces provide you to deploy for demo and it will be for free you can just try out there okay in a sense there will be some limited token to so you can easily for learning phase it is enough to learn okay you can even use streamlit it's a library you can install and you can directly run over it there is something called serverless deployment also you can use either Azure AWS or gcp in Azure you have something called Azure AI studio and Azure ml where you can you can configure the things clearly and in AWS you have Bedrock where everything without any server it can work things clearly even you can configure the server also it's of your choice okay and Edge deployment you can learn about mlc and llm and MN and llm these two Frameworks will help you to perform Edge deployment inance like you can see or have already return here yeah so which are nothing but you will be deploying them in a web browser or an Android or an iOS so for that you can use them okay and there is an optional topic also called securing llm which is nothing but prompt hacking back door and defensive measures so if you're building a website obviously you should even check the security of the TMS or even the website right so then you can use them Lang fuse grag is there you can explore about them in detail okay so it was a very long video guys I hope you all like this particular video and the explanation you will be getting this uh road map in this video description link you can just open that and everyone can see it and you can even learn generate UI okay so guys that's all for today's video do like share and subscribe the channel thank you all meet you in the next video

Channel: SAI KUMAR REDDY

Views: 2,657

Keywords: data science, artificial intelligence, machine learning, NLP, Roadmap for Beginners, future of nlp, nlp interview questions, nlp for text analysis, data science for beginners, data science tutorial, Transformers, encoder and decoder, ai/ml roadmap, ai/ml tutorial, gpt 4, gpt, bert, lama 2, what is data science, ai, langchain, mistral ai, mistral 7b, gemini, llama2, lamborghini, sai kumar reddy, genai, generative ai, generative, ai roadmap, ai roadmap for beginners

Id: QHIOC2UcHZ8

Length: 34min 42sec (2082 seconds)

Published: Mon Mar 11 2024