- 9 months ago
Dive into the Fundamentals of Generative AI in this insightful lecture from the AI-900 Microsoft Azure AI Fundamentals course. Discover how Generative AI is transforming industries by creating new content, automating processes, and enhancing creativity.
What you’ll learn:
What is Generative AI, and how does it work?
Key techniques: Deep Learning, GANs (Generative Adversarial Networks), and Transformers
Applications of Generative AI: text generation, image creation, and data augmentation
Real-world use cases of Generative AI in industries like media, healthcare, and technology
Introduction to Azure’s capabilities for Generative AI
🎯 Perfect For:
Beginners in AI & Machine Learning
IT Professionals
Students & Developers
Azure Certification aspirants
📚 Explore Our Other Azure Courses on: https://www.youtube.com/@skilltechclub
What you’ll learn:
What is Generative AI, and how does it work?
Key techniques: Deep Learning, GANs (Generative Adversarial Networks), and Transformers
Applications of Generative AI: text generation, image creation, and data augmentation
Real-world use cases of Generative AI in industries like media, healthcare, and technology
Introduction to Azure’s capabilities for Generative AI
🎯 Perfect For:
Beginners in AI & Machine Learning
IT Professionals
Students & Developers
Azure Certification aspirants
📚 Explore Our Other Azure Courses on: https://www.youtube.com/@skilltechclub
Category
🤖
TechTranscript
00:00okay now we are at the last module of this particular session which is
00:12Microsoft Azure AI fundamentals focusing on generative AI or in short gen AI is
00:18one of the most popular module of recent time and obviously chat GPT is
00:24something which is a very famous word now which is a part of generative AI so
00:29let's focus on this module in this module first we are going to focus on
00:33fundamentals of generative AI what kind of things we need to understand inside
00:37this and then fundamentals of Azure open AI service now as we know open AI is a
00:43separate company and they are the one who has invented something which is known
00:48as chat GPT now open AI is separately available as a service from open AI dot
00:54org as well as we have an collaboration of open AI with Azure which is known
00:59as Azure open AI service in this module we'll see how we can provision that
01:03service what kind of things we have to keep in mind and how we will use it at
01:08the end of this we have to explore responsible generative AI principles now
01:12we know we have a responsible AI principles which we have seen in the
01:16first module of this course same way while using generative AI also there are
01:21certain guidelines which we have to keep in mind those things we are going to see
01:24in this module the learning objectives are going to focus on first what is
01:30actually generative AI and how it actually works what kind of architecture it
01:35follows is all things which we are going to see in the first part in the
01:39second step we are going to focus on describing the capabilities of large
01:43language model formerly known as LLM this LLMs are those models which are
01:48created by organizations like open AI and this models are actually providing the
01:55the knowledge base kind of association for your generative content you can
02:01generate text you can generate images you can actually try to understand and
02:06analyze text based on that there are many capabilities generative AI is having
02:11all this are because of this large language models and then we are going to
02:15focus on the third one which is understand how to use a sure open AI to
02:20create a generative AI solutions we'll focus on how we can create this and how
02:25we can configure a sure open AI service in this module now let's start with
02:31fundamentals of generative AI definition wise what is generative AI now we
02:37already know AI is something which is going to imitate human behavior by using
02:42machine learning to interact with the environment and execute tasks without
02:47explicit directions on what to output now this is a simple common definition of
02:53AI which we have seen in the first module also which is like it's going to
02:57mimic the human behavior while in the other case we have generative AI which is
03:02going to create an original content remember using natural language
03:06generation image generation and code generation whatever things which are
03:11generated by generative AI all will be an original content now this generated
03:18content is going to be not going to have any copyright association with that and
03:22it's not a copy of any existing content available on internet generative AI
03:26application can take this kind of a natural language input maybe from the text or
03:32speech and then it can return the appropriate responses in variety of formats
03:36like it is mentioned here now because we need to understand generative AI we need to
03:42understand how it is working and the first step in that is large language models
03:46generative AI applications are powered by LLM and these are a specialized type of
03:54machine learning models that you can use to perform natural language processing task which
04:00includes like determining the sentiment or classification of the natural language text it can
04:06help you to summarize the text even the lengthy documents it can help you to compare multiple
04:11text sources for semantic similarities and then it can give you what similarities or
04:16differences are there and it can also help you to generate a new natural language
04:21where a totally fresh new content will be generated based on that now when we have LLM large language
04:27models the next question which comes arises how it is working what kind of things are
04:33actually generating this kind of a new content well the answer of that is this we have a model called
04:40transformer model for your kind of information if you already know GPT the full form of GPT is
04:47generative pre-trained transformer now in this case the word transformer is actually representing to this
04:54model which is known as transformer this is the core architecture which consists of two main parts
05:00there are two parts which are here you can see there is one thing called encoder and there's another
05:05thing called decoder the encoder block that creates semantic representation of a training vocabulary
05:12so whatever input which you are going to provide inside this is going to use the encoder to analyze and
05:18understand that with the existing vocabulary which are available with that then the decoder block is
05:24going to generate the new language sequences let's say you're going to provide some kind of a text input
05:30which is one statement in english language when you're going to provide this the first thing
05:36which is going to happen is is going to tokenize this each word of this particular phrase that token
05:43is nothing but a numeric token which is going to represent a unique number associated with that
05:49the tokens are assigned in embedding and then this is going to be converting that into a vector values
05:55with multiple dimensions now if you don't know what is the tokenization i strongly recommend you to go to
06:01an openai website link which is this platform.openei.com slash tokenizer when you go there and if you just
06:10put some particular text there it's going to show you how gpt 3.5 4 and gpt 3 which was a legacy version of
06:18that is going to tokenized any content let's say i'm going to put some text here like
06:23i am learning azure openai and it is really useful for my day-to-day work now when i put this this is
06:46one particular statement but if you see at the below this is showing me that this statement is containing
06:51total 18 tokens each token is going to be associated with the multiple characters which are used in
06:58each word now when i want to see these tokens we have a text which is here and if i click on the token
07:05id is generating some particular numeric values associated with that each value associated with
07:12that word is going to be unique sometimes it's also going to count a group of words into that
07:17now this whole configuration is something which is known as tokenizer and every time when you have
07:23provided any text in generative transformer model is going to first tokenize that and that's what that
07:30first step is actually all about after tokenization we have something which is known as attention now if
07:37you see this transformer model you need to understand this transformer model from the bottom to
07:42this particular flow so the flow is going to be starting from this input text then it's going to
07:48this way and then it's going to be giving you the output at the end of this so this is how we have to
07:52understand that the first step is input text which is going to have a tokenization associated with that
07:58and after that the embedding is going to start which is going to have the vector generated with that
08:03once the vector is there inside the encoder block initially the vocabilities are going to be associated
08:09with this vector and then in the decoder block is going to generate the content that is what actually
08:14going to happen but now inside encoder and decoder block we have multiple attentions which are
08:20available here you can see each part of this is actually nothing but some kind of an attention
08:25what is an attention the attention layer is going to examine each token in turn and then determine the
08:31embedding values that reflect the semantic relationship between the tokens in simple words when you have
08:38multiple tokens associated with each other obviously in a statement multiple words are going to be
08:44related with each other because ultimately a statement is nothing but sequence of multiple words
08:49how one word is connected with other one that's going to define the meaning of the statement and that's
08:55exactly same thing what attention is doing while taking the input and while generating the output
09:02words should be organized in the proper order then only this is going to be coming to a proper
09:07meaningful statement and that's what this attention blocks are actually doing in the decoder side the
09:13relationships are used to predict the most probable sequence of the token so sequencing in the input
09:18is going to decide the sequencing in the output that's how the generative transformer model works
09:24now going deep dive into transformer model is something which is required if you are going to work
09:30on open ai or any other generative transformer models but as a part of this particular course i think that
09:37is something which is out of the scope so we are not going into that depth but yes this words tokenization
09:43embeddings encoder decoder block and attention if you understand these terms then this is something
09:49which is more than enough for this particular module this slide is showing you how tokenization is going to
09:55work as i have shown you in that website once you put a statement like here we have an example sentence
10:01which is i heard a dog bark loudly at a cat now when i see this here each word is going to be generated
10:10with one numeric value here we are just having a values like one two three four five six seven eight but
10:15in real tokenization it's going to have a unique id associated with that each word as i have shown you in
10:20that website link then that is what tokenization is going to happen after this they are going to
10:26convert that into a vector which is going to be a you can say three dimensional object associated with
10:32that and that is what which is going to happen in embeddings so the relationship between these tokens
10:37are captured as vectors and these are something which are known as embeddings that's the definition of
10:42embeddings in short it's going to think like a three-dimensional box kind of thing each word which is a dog or a cat
10:49or a bark is all going to happen and associate with this kind of a vector once you have the vectors
10:55it's going to associate that okay if we have a bark that is how it is going to be associated with the dog
11:00and if we have a cat the cat is going to sound like meow something like that the relationship between
11:05word is going to happen finally after this we have a third stage which is attention which is happening
11:11in encoder and decoder both the blocks this is going to capture the strength of the relationship between
11:16the tokens using the attention technique this is a one of the useful technique of transformer model
11:22where it's going to pass on this vector graphics into this particular neural network and those neural
11:29networks are going to process this information finding the relationship between those words and then
11:34generating the content based on that the repetitive process of this is going to happen multiple times
11:39and that is what which is going to take care of the processing of generative content
11:44copilons are often integrated into your existing applications and it's going to provide a way
11:50better for users to get help with the common task of the generative ai model now it can help you to
11:57integrate with your existing applications like microsoft office or maybe your microsoft edge browser
12:04or even with the windows operating system the best way to use it is maybe in the bing search engine
12:09where while searching it you can use the generative ai associations with that using these developers
12:15can build copilot that submit prompts to a large language models and generate content for the use in
12:21application business users can use copilot to boost their productivity save their time while generating
12:27the content and making sure that their creativity with ai generated content is going to be smoothly working
12:34with that now as you can see here in the slide in the microsoft powerpoint we are able to generate
12:39images with the help of copilot if you integrate it with that what are the other options as i said we
12:46have edge browser available with copilot we have bing copilot also while searching as well as we have
12:52github copilot which can generate a code content for you in various programming languages which can help
12:58you and save you a lot of time with that i strongly recommend you to try some of the microsoft learn
13:03labs of copilots if you're really interested in that now let's talk about improved generative ai
13:09responses with prompt engineering the quality of responses that generative ai application returns
13:16not only depends on the large language model itself it is actually something which is a type of the prompt
13:22which you have given and that is where the term prompt engineering is coming into the picture
13:28this is again one of the popular term both developers who design the application and consumers
13:34who use the application can improve the quality of responses from generative ai by using a direct
13:41language you can provide system messages you can provide the input in the text format or you can
13:46provide something which is some kind of an example known as few short or one short kind of a thing
13:52and using that you can improvise your response which you are getting from the generative ai now this
13:59prompts are the way by which you can tell the application what you actually want this application
14:04to do and when you're providing this configuration writing a good prompt is going to help you to achieve
14:11a desired better result when you submit a clear and a specific prompts with that like you can see here we
14:18have a table which is showing you a description and the example of that particular prompt engineering
14:24method now you can use a direct language like in the description is saying that you can get most of
14:30the useful completions by being explicit about the kind of response you want so while specifying in the
14:36direct language you can say create a list of 10 things to do in edinburgh during august now when you say
14:43create a list it is obviously going to give you a bulleted list kind of a thing so you're directly specifying
14:48in the language that what exactly you're looking for same way you can see we have a configuration of
14:52system message in most of the chat bot kind of a configuration in the system message you can specify
14:58how that chat bot is going to behave and act so you can see you are specifying you are a helpful
15:04assistant that responds in a cheerful and friendly manner so what kind of assistant is this how it is
15:10going to response all this we have to specify inside the system message configuration very soon i'm going
15:16to show you an azure ai demo in which we are going to use chat bot with some kind of a customized system
15:22messages with that providing examples you can specify multiple examples like how you want to use that
15:31like this is something which i say that like few short one short kind of thing so you can visit the
15:36castle in the morning before the crowds arrive now this is one particular example which we are providing
15:42and then this is something which is known as few short learning so while giving the question you're
15:47also suggesting that this kind of answer we are expecting and we are giving some example of that
15:52and finally the fourth one is we have grounding data so we can include our grounding data to provide
15:58the context like we can include an email text with the prompt which is like summarize my email so you're
16:04already providing an email and then you're just specifying summarize this email with that so you're already
16:09providing grounding data and based on that you're expecting some kind of a response this is what
16:14prompt engineering is all about but again this is also one of the topic which is super super huge you
16:21can spend days if not months if you want to deep dive into prompt engineering right now let's understand
16:29this prompt engineering and configuration properly uh in our azure open ai service but yes there is one
16:36exercise which is available on microsoft learn if you have a time you can try the generative ai with
16:42bing chat copilot now you know bing is a search engine from microsoft you can use bing search engine
16:50and inside that you can configure the copilot and then you can generate a content using that that is what
16:55this particular exercise is actually trying to show you i strongly recommend you to go with that
17:01but in our case we are directly going to use azure open ai service now
Comments