🪧

Generative AI course

1. What is Generative AI?

2. Roadmap of Generative AI?

3. What things do we need to learn in a Generative AI course?

4. What are the foundational models in Generative AI?

5. What is the future scope of Generative AI?

6. What is the eligibility for a Generative AI course?

7. Why should you learn Generative AI?

8. What are the types of Generative AI compared to other AI types?

9. What are the advantages and disadvantages of Generative AI?

Generative AI course 1
 what is Generative AI
Generative AI refers to AI technologies that can create new content, ideas, or
data that are coherent and plausible, often resembling human-generated
outputs. It has a plethora of practical applications in different domains such as
computer vision, natural language processing, and music generation.

Roadmap for Generative AI

1. Foundation in Machine Learning and Deep Learning

Ensure proficiency in probability, statistics, linear algebra, and calculus.

Gain hands-on experience with programming languages like Python/R.

Familiarize yourself with supervised and unsupervised learning algorithms.

Build machine learning models on tabular datasets.

2. Deep Learning Mastery

Develop a solid understanding of deep learning architectures such as
MLPs, RNNs, LSTMs, GRUs, and CNNs.

Gain proficiency in at least one deep learning framework like Keras,

TensorFlow, PyTorch, or FastAPI.

Train MLPs on tabular datasets.

Construct RNNs and CNNs for unstructured data (text and image).

3. Advanced Deep Learning Techniques

Learn about pretrained models for image data and their types.

Understand language models and build them using LSTMs/GRUs.

Explore attention mechanisms and their applications.

Study autoencoders and GANs architectures and train these models on

datasets.

4. Generative Models for NLP:

Discover Large Language Models (LLMs) like Transformers, BERT, GPT,
PaLM, etc.

Generative AI course 2
 Understand how to use LLMs for downstream tasks: finetuning, zero-shot,
one-shot, and few-shot learning.

Learn best practices for training LLMs, including scaling laws and efficient
training mechanisms.

Explore techniques to pre-train LLMs on domain-specific data.

Implement different techniques to finetune LLMs for downstream tasks.

Study optimization techniques like Adapters, LoRA, QLoRA, etc., to

accelerate finetuning.

Understand deployment considerations (LLMops) for deploying LLMs in

production.

Explore cutting-edge models like ChatGPT and BARD and understand their
training process, including reinforcement learning from human feedback,
supervised fine-tuning, and prompt engineering.

Gain hands-on experience with LLM frameworks like LangChain, AutoGPT,

Vector DB etc.

Disadvantages :

1. Source Ambiguity: Generative AI may not always attribute content to its

original source, leading to ambiguity in authorship and ownership.

2. Bias Assessment Difficulty: Assessing bias in original sources can pose

challenges, complicating efforts to ensure fairness and accuracy in
generated content.

3. Realism Challenges: The lifelike quality of generated content can obscure

inaccuracies, making it more difficult to distinguish between genuine and

Generative AI course 3
 fabricated information.

4. Tuning Complexity: Adapting generative AI for new contexts and scenarios

can be intricate, requiring nuanced adjustments to ensure optimal
performance.

5. Potential for Oversights: Despite advancements, there remains a risk that

generative AI results may inadvertently perpetuate biases, prejudices, and
hateful ideologies, warranting careful scrutiny and mitigation measures .

What are the benefits of generative AI?

Automating the manual process of writing content.

Reducing the effort of responding to emails.

Improving the response to specific technical queries.

Creating realistic representations of people.

Summarizing complex information into a coherent narrative.

Simplifying the process of creating content in a particular style.

What are the concerns surrounding

generative AI?

It can provide inaccurate and misleading information.

It is more difficult to trust without knowing the source and provenance of

information.

It can promote new kinds of plagiarism that ignore the rights of content
creators and artists of original content.

It might disrupt existing business models built around search engine

optimization and advertising.

Generative AI course 4
 It makes it easier to generate fake news.

It makes it easier to claim that real photographic evidence of a wrongdoing

was just an AI-generated fake.

It could impersonate people for more effective social engineering cyber

attacks.

What are some examples of generative AI

tools?

Text generation tools include GPT, Jasper, AI-Writer, and Lex.

Image generation tools include DALL-E 2, Midjourney, and Stable

Diffusion.

Music generation tools include Amper, Dadabots, and MuseNet.

Code generation tools include CodeStarter, Codex, GitHub Copilot, and

Tabnine.

Voice synthesis tools include Descript, Listnr, and Podcast.ai.

AI chip design tool companies include Synopsys, Cadence, Google, and

Nvidia.

What are use cases for generative AI?

Implementing chatbots for customer service and technical support.

Deploying deepfakes for mimicking people or even specific individuals.

Generative AI course 5
 Improving dubbing for movies and educational content in different
languages.

Writing email responses, dating profiles, resumes and term papers.

Creating photorealistic art in a particular style.

Improving product demonstration videos.

Suggesting new drug compounds to test.

Designing physical products and buildings.

Optimizing new chip designs.

Writing music in a specific style or tone.

Some of the very commonly known LLMs

are:

Open AI’s GPT 3, 3.5, and 4

Google’s LaMDA and PaLM

Hugging Face’s BLOOM

Meta’s LLaMA

NVidia’s NeMO LLM

Types of Models for Generative AI ?

1. Generative Adversarial Networks (GANs): GANs consist of two neural

networks, a generator and a discriminator, trained simultaneously in a
game-like scenario. The generator tries to create data (e.g., images) that
are indistinguishable from real data, while the discriminator tries to

Generative AI course 6
 differentiate between real and generated data. This adversarial training
process helps improve the quality of generated samples.

2.

1. Variational Autoencoders (VAEs): VAEs are generative models based on

the principles of autoencoders. They consist of an encoder network that
compresses input data into a latent space representation and a decoder
network that reconstructs the original data from the latent space. VAEs are
trained to generate new data samples by sampling from the learned latent
space distribution.

2.

1. Transformers: Transformers are a type of deep learning model primarily

used in natural language processing (NLP). They are based on a self-
attention mechanism that allows the model to weigh the importance of
different words in a sequence when processing each word. Transformers
have achieved state-of-the-art performance in various NLP tasks and have
also been adapted for other tasks beyond NLP, such as image generation
and reinforcement learning.

2.

Generative AI course 7
 1. Autoencoders: Autoencoders are neural networks trained to copy their
input to their output. They consist of an encoder network that compresses

Generative AI course 8
 the input data into a latent space representation and a decoder network that
reconstructs the original input from the latent space representation.
Variants like denoising autoencoders, sparse autoencoders, and contractive
autoencoders serve different purposes.

2.

Future capabilities of generative AI:

Continued Model Scaling and Efficiency Gains
Larger and more parameter-dense models promise to further expand
generative capabilities and general competencies.

At the same time, more efficient training techniques will reduce data and
computing needs to improve sustainability.

Also, new hardware and software architectures could be developed to

more effectively handle resource intensive workloads.

Generative AI course 9
 Steady Improvements in Output Quality and
Accuracy
As this technology advances, its responses may become indistinguishable
from, and possibly superior to human-generated material.

Improvements will be achieved through advances in responsible

development, model architecture, training techniques, data quality and
diversity, and evaluation metrics and tools.

Also, end users and developers contribute to improvements by providing

feedback on responses, labeling data, generating ideas, developing
models, and building tools.

Multiple Layered Agent Frameworks:

Multiple layered agent frameworks are a complex and hierarchical
structure of AI agents designed to interact and collaborate effectively in a
dynamic environment.

This type of framework allows numerous agents to communicate with one

another in order to solve problems that are too complicated for a single
agent to accurately accomplish on its own.

In addition, multiple AI agents working together can dramatically increase

the productivity of task completion.

For instance, picture a transportation network management system where

sensor agents gather real-time data, data processing agents analyze it,
routing agents determine optimal paths, coordination agents ensure
harmonious cooperation, and supervisory agents oversee the entire
operation.

Autonomous AI Agents
One or more AI agents with the ability to learn and adapt on their own,
without the need for human intervention.

This will allow them to automate tasks and make decisions in real time.

For example, an autonomous AI agent could be used to monitor current

events and automatically generate content on flagged topic categories.

Generative AI course 10
 Increased Customization and Specialization :
Generative AI will become more customizable, allowing for more
parameters to be fine-tuned on custom data.

This contextual fine-tuning improves relevance, quality and safety for

specific professional use cases while retaining general advantages.

Which will serve as springboards for developing specialized

systems tuned on custom datasets for focused tasks, verticals and
applications.

To illustrate, a healthcare company could fine-tune a generative AI model

on medical data to generate new drug candidates, predict the effectiveness
of different drugs, and design new drug compounds.

Accessibility :
Generative AI tools will become more accessible to people with less data
and expertise.

Making it easier for businesses of all sizes to use this technology by

lowering barriers to access, enabling broader adoption across languages,
domains, geographies and less resourced organizations.

Specifically, the development of user-friendly apps and AI-as-a-service

(ALaaS) platforms will enable access without having to build and maintain
their own infrastructure.

Text to Everything
The ability to generate text, images, audio, videos, and 3D environments
from a single platform.

This means users will provide a single text prompt and the model will
generate every content format simultaneously.

For example, a user could provide a prompt like “A beautiful sunset over a
beach” and the model would generate a text description of the sunset, a
video of the sunset, and an audio file of the sound of waves crashing on the
shore.

Generative AI course 11
 Ultra Realistic Talking Avatars :
Generative AI models will be able to create videos of talking avatars that
are so realistic that they are indistinguishable from real people.

These avatars will be able to lip sync to speech, have natural facial
expressions, subtle head movements and even generate their own
dialogue.

This will have a wide range of applications, including virtual assistants,

customer service representatives, and even educational characters

Impact of Generative AI on industry and

jobs :

Changing Job Roles and Required Skills :

Though AI won’t replace human roles outright, it will alter the associated
required skills.

workers and job requirements will need to transition into more strategic
roles.

This may involve upskilling to perform higher-value work like data

analysis, creative direction and strategy.

Automating Tasks to Increase Efficiency

Conducting process analysis can be used to identify repetitive and low
strategic value tasks that are ideal for AI automation.

Task categories such as document review, contract analysis, customer

service routines and content generation contain repetitiveness that humans
find tedious.

Generative AI course 12
 free up employees for more impactful and creative responsibilities better
suited to human strengths.

Renewed Focus on Data Governance

Generating high-quality outputs relies on extensive training data that is
accurate, comprehensive and aligned to use cases.

This places greater emphasis on responsible data governance, including

ethics, integrity, monitoring, security and compliance.

Where data collection practices will require more comprehensive

auditing to avoid perpetuating biases.

Tight Integration into Workflows and Software

To maximize user adoption and impact, AI systems will integrate
seamlessly into existing enterprise software, workflows and processes.

This integration will emphasize on providing a high-quality user

experience by actively seeking continuous feedback to drive
enhancements.

Ultimately, the main objective of these AI systems is to seamlessly

integrate into workflows, to feel like a natural system component rather
than function as isolated tools.

Lower Barriers of Entry

Generative AI will make it easier for new competitors to enter the market.

As pre-trained models and AI-as-a-service (ALaaS) platforms become more

accessible, smaller players will leverage these technologies without
massive in-house investments, somewhat leveling the playing field.

As a result, this will increase market competition and force existing

enterprises to adapt quickly and differentiate themselves.

Enabling New Niche Capabilities

With the help of generative AI, one can scan the market for untapped
areas to uncover potential niche opportunities.

Generative AI course 13
 Specifically, to identify new market opportunities from various sources by
analyzing customer data, social media, demographics, economic trends,
etc.

By feeding all these data points together, the AI model will be able to
identify new customer needs and wants that are not currently being met
by existing products and services.
• Then
integrate such findings into the process to generate new ideas for
products and services.

Applications Across Industries:

Marketing, Advertising, and Entertainment Industry:

1. Content Creation: Generative AI powers content creation in various

forms like art, music, and literature, allowing artists and musicians to
explore innovative creative directions.

2. Video Game Development: AI-driven systems create game

environments, characters, and dialogues, reducing development time
and resources.

3. Scriptwriting: Screenwriters use Generative AI to assist in scriptwriting

by generating dialogues, plotlines, and character interactions.

Education Sector:

1. Personalized Learning: Generative AI adapts educational content to

individual student needs by generating tailored assignments, quizzes,
and study materials.

2. Knowledge Base: AI creates exhaustive knowledge bases for students

to access information in a conversational style.

3. Virtual Labs: Generative AI powers virtual laboratories for simulating

experiments and scenarios in science and engineering disciplines.

Healthcare Industry:

1. Medical Image Generation: Generative AI creates synthetic medical

images for training models, enhancing diagnostic accuracy, and
simulating rare medical conditions.

Generative AI course 14
 2. Drug Discovery: Pharmaceutical companies use Generative AI to
discover new drug compounds by generating molecular structures.

3. Personalized Medicine: AI analyzes patient data to generate

personalized treatment plans, accounting for genetic factors and
medical history.

Manufacturing Industry:

1. Product Design: Generative design creates optimized product designs

considering materials, weight, and structural integrity.

2. Quality Control: AI generates synthetic data for quality control testing,

ensuring adherence to quality standards.

3. Supply Chain Optimization: AI-generated forecasts help make informed

decisions about production and distribution.

Software & Tech Industry:

1. Code Generation: Generative AI assists developers by generating code

snippets and templates, speeding up development.

2. Bug Detection: AI generates synthetic test cases to identify and fix

software bugs efficiently.

3. IT Security: Generative AI models simulate cyberattack scenarios to

identify vulnerabilities and enhance cybersecurity measures.

What is prompt engineering

Generative AI course 15
 Prompt engineering is the practice of designing inputs for AI tools that will
produce optimal output

why we need to learn prompt enginerring ?

Enhanced accuracy and relevance

Improved decision making

Personalized customer experiences
Efficient resource utilization
Ethical considerations and bias mitigation

There are 8 prompt engineering methods:

(1) Zero-Shot Learning, (2) One-Shot Learning, (3) Few-Shot Learning,

(4) Chain-of-Thought Prompting, (5) Iterative Prompting,
(6) Negative Prompting, (7) Hybrid Prompting, and
(8) Prompt Chaining .

1.
Zero-Shot Learning: This involves giving the AI a task without any prior
examples. You describe what you want in detail, assuming the AI has no prior
knowledge of the task.

Generative AI course 16
 . One-Shot Learning: You provide one example along with your prompt. This
helps the AI understand the context or format you’re expecting.
3. Few-Shot Learning: This involves providing a few examples (usually 2–5) to
help the AI understand the pattern or style of the response you’re looking for.
4. Chain-of-Thought Prompting: Here, you ask the AI to detail its thought
process step-by-step. This is particularly useful for complex reasoning tasks.

5. Iterative Prompting: This is a process where you refine your prompt based
on the outputs you get, slowly guiding the AI to the desired answer or style of
answer.
6. Negative Prompting: In this method, you tell the AI what not to do. For
instance, you might specify that you don’t want a certain type of content in the
response.

7. Hybrid Prompting: Combining different methods, like few-shot with chain-of-

thought, to get more precise or creative outputs.
8. Prompt Chaining: Breaking down a complex task into smaller prompts and
then chaining the outputs together to form a final response.

Key elements of a prompt :

Instruction. This is the core directive of the prompt. It tells the model what
you want it to do. For example, "Summarize the following text" provides a
clear action for the model.

Context. Context provides additional information that helps the model

understand the broader scenario or background. For instance, "Considering
the economic downturn, provide investment advice" gives the model a
backdrop against which to frame its response.

Input data. This is the specific information or data you want the model to
process. It could be a paragraph, a set of numbers, or even a single word.

Output indicator. Especially useful in role-playing scenarios, this element

guides the model on the format or type of response desired. For instance,
"In the style of Shakespeare, rewrite the following sentence" gives the
model a stylistic direction.

Generative AI course 17
 Technical skills for prompt engineering :

Understanding of NLP. A deep knowledge of Natural Language Processing

techniques and algorithms is essential.

Familiarity with LLMs. Experience with models like GPT, PaLM2, and other
emerging models their underlying architectures.

Experimentation and iteration. Ability to test, refine, and optimize prompts

based on model outputs.

Non-technical skills for prompt engineering

Communication. The ability to convey ideas, collaborate with teams, and

understand user needs.

Subject Matter Expertise. Depending on the application, domain-specific

knowledge can be invaluable.

Language Proficiency. Mastery over language, grammar, and semantics to

craft effective prompts.

Critical Thinking. Evaluating model outputs, identifying biases, and

ensuring ethical AI practices.

Creativity. Thinking outside the box, experimenting with new prompt styles,
and innovating solutions.

Elements of a Prompt

Generative AI course 18
 A prompt can include different elements:

Instruction — It’s a specific task or direction for the model to follow. :

Context— It provides extra information or context to help the model

generate better responses.

Input Data — It refers to the question or input for which we want the model
to provide a response.

Output Indicator — It indicates the desired type or format of the output.

Prompt Engineering use cases:

1. Text summarization

2. Information Extraction

3. Question Answering

4. Text Classification

5. Code generation

6. Chatbots and Virtual Assistants

Effective prompt engineering involves several best practices:

1. Clearly Define Desired Responses: Specify the scope of the desired

response to prevent misinterpretations by LLMs.

2. Be Specific and Explicit: Avoid vagueness; provide clear, explicit

instructions to guide LLMs effectively.

3. Balance Simplicity and Complexity: Find the right balance between simple
and complex prompts to ensure clarity without overwhelming Large
Language Models.

4. Iterate and Experiment: Prompt engineering is iterative; test different

prompts and refine them based on results.

5. Train and Evaluate: Consider additional training to improve AI model

performance if necessary, creating custom models for specific tasks.

Generative AI course 19
 Task
“Task” is the actual prompt, but without the constraints.

Context
“Context” is the actual detailed content and it is the 2nd most important and
essential part of your prompt

Examples
“Examples” is also an important part of proompting since giving clear and
concise examples tells the large language model about the desired outcome
and style

Persona

Generative AI course 20
 “Persona” is the moment when you say: “Write Like You Are” or “Create in
the style of” to create the content as if that person is writing it.

Format
“Format” is about in what shape or form you want your results to be.

Tone
“Tone” is when you say: professional, friendly, casual, concise, emotional, etc.
It is the feeling that you want to convey to your readers.

Prompting Principle :

Generative AI course 21
 Principle 1: Write clear and specific instructions

Tactics

1: Use delimiters to clearly indicate distinct parts of the input

Delimiters can be anything like: ```, """, < >, <tag> </tag> , :

2. Ask for the strutured output

3. ask model to check whether condition satisfied or not

4. few shot prompting

Principle 2: Give the model time to “think”:

1: Specify the steps required to complete a task

2. : Instruct the model to work out its own solution before rushing to a
conclusion

Model Limitations: Hallucinations

How to Iterate your prompt for getting the actual

output from it :
1. The text is too long
2.Text focuses on the wrong details
3.description in not as you wanted

2. Summarizing :

Generative AI course 22
 1.Summarize with a word/sentence/character limit
2.summarize with your focus point like price , values

3. also you can summarize list of review by prompting

3. Inferring :
1. Entity Recognition: Identifying and categorizing entities mentioned in text,
such as people, organizations, locations, dates, and more.

2. Part-of-Speech Tagging: Assigning grammatical categories (e.g., noun,

verb, adjective) to words in a sentence.

3. Named Entity Recognition (NER): Identifying proper nouns that refer to

specific entities, such as names of people, places, organizations, etc.

4. Coreference Resolution: Determining which words or phrases in a text refer

to the same entity.

5. Semantic Role Labeling: Identifying the roles of words or phrases in relation

to a predicate (e.g., who did what to whom).

6. Textual Entailment: Determining the logical relationship between two pieces

of text, such as whether one sentence entails or contradicts another.

7. Question Answering: Generating accurate responses to questions posed in

natural language based on a given context.

sentiment analysis
identify the emotion
extract multiple things

4. Transforming :

Generative AI course 23
 language correction
language translation
Tone transformation
format coversion

spell and grammer checker

5. Expanding : convert short information

into longer one:
replying anyone

writing essay
explaining topic

6. Build your Bot :

Generative ai Terms and Terminolgy :

LIFE CYCLE OF GENERATIVE AI PRODUCT

. Objective: Define Business Use Case

Generative AI course 24
 In this initial phase, the primary focus is on articulating the specific business
challenge where leveraging a Large Language Model (LLM) strategically for
enterprise applications is crucial. Key selection criteria include:

Identifying a suitable use case that aligns with business objectives.

Ensuring the capabilities of the chosen LLM align with the identified
challenge.

Assessing execution feasibility in terms of available resources and required

skills.

Hypothesizing potential business value generated from implementing the

LLM solution.

2. Select Model: Choose the Right LLM Model for the Business Case

This stage involves strategically selecting the most suitable LLM model tailored
to meet the unique demands of the identified business use case.
Considerations include:

Choosing between models such as GPT-3 or LLMA for text generation,

based on specific requirements.

Evaluating models specialized in tasks like Sentiment Analysis or

multilingual translation, such as M2M-100.

Aligning the choice with the production or research environment and

scalability needs.

3. Adapt and Harmonize: Prompt Engineering

The focus here is on crafting input prompts strategically to guide LLMs in

generating precise and relevant outputs. Benefits include:

Generating accurate and contextually relevant responses to queries.

Providing contextual guidance to mitigate biases and improve response

quality.

Saving time and resources by streamlining LLM output.

4. Adapt and Harmonize: Contextual Enrichment with RAG

Generative AI course 25
 Introducing Retrieval Augmented Generation (RAG) enriches LLM performance
by providing additional context during text generation, leading to:

Enhanced accuracy, depth, and relevance of generated outputs.

Reduction of inaccuracies or hallucinations in responses.

Improved overall LLM performance.

5. Adapt and Harmonize: Model Refinement through Instruction Fine-tuning

Instruction fine-tuning involves supervised learning to enhance LLM
performance across different tasks by:

Training the model with labeled examples to improve task-specific

performance.

Utilizing techniques like LoRA fine-tuning or soft prompts to adjust model

behavior.

Enhancing model adaptability and performance through iterative

refinement.

6. Adapt and Harmonize: User-Centric Improvement via Feedback Integration

Iteratively incorporating user feedback ensures that LLMs align with user
expectations and ethical standards by:

Establishing a continuous improvement cycle based on user feedback.

Addressing biases and privacy concerns to enhance user trust.

Ensuring LLM development aligns with user requirements and ethical

guidelines.

7. Adapt and Harmonize: Comprehensive Model Assessment with LLM

Performance Evaluation
Assessing LLM performance involves employing specialized metrics like
ROUGE and BLEU to:

Evaluate generated outputs in tasks such as summarization and translation.

Gain nuanced insights into model efficacy while acknowledging evaluation

complexities.

Generative AI course 26
 Utilize multiple benchmarks like GLUE, SuperGLUE, MMLU, BIG-Bench, and
HELM for comprehensive assessment.

8. Application Integration: Optimize and Deploy LLMs

Transitioning from training to real-world application involves optimizing LLMs
for performance and resource efficiency through techniques such as:

Model distillation to train smaller models mimicking larger ones.

Quantization to reduce memory usage and improve efficiency.

Pruning to streamline the model without sacrificing accuracy.

Deploying optimized LLMs can occur on-site, on the cloud, or on edge devices
as per requirements.

9. Application Integration: Develop Applications Empowered by LLMs

In this final stage, applications are tailored and refined using LLMs to ensure
precise responses by:

Integrating new data and considering domain-specific contexts.

Offering accurate answers for tasks ranging from chatbots to decision-

making systems.

Fine-tuning models based on user input to enhance performance and

relevance.

Generative AI course 27
 1. Data Preparation: Data Preprocessing and Filtering
In this stage, the focus is on preparing the data for training by performing
preprocessing and filtering. This involves:

Cleaning and organizing the data to remove noise and inconsistencies.

Preprocessing text data by tokenizing, normalizing, and removing stop

words or irrelevant information.

Filtering the data to ensure it meets quality standards and is relevant to the
model's objectives.

2. Train Model: Select a Foundational Model and Begin Training

Here, the foundational model is selected based on the task requirements, and
training is initiated by providing the model with tokenized data. Steps include:

Choosing an appropriate foundational model, considering factors like

architecture, size, and pre-trained weights.

Providing tokenized input data to the model to start the training process.

Monitoring and adjusting training parameters to optimize model

performance and convergence.

3. Validate the Model: Create Model Card

Validation involves creating a model card to provide comprehensive information

about the trained model. Key components include:

Generative AI course 28
 Documenting model architecture, parameters, and training data.

Describing model performance metrics and evaluation methods.

Disclosing any biases, limitations, or ethical considerations associated with

the model.

4. Tuning the Model: Fine-Tuning for Optimization

Fine-tuning the model aims to optimize its performance further based on
validation results. Steps include:

Adjusting model parameters, such as learning rate or batch size, based on

validation feedback.

Fine-tuning the model on specific tasks or domains to improve accuracy

and generalization.

Iteratively refining the model to achieve desired performance levels.

5. Deploy the Model:

Deployment involves preparing the model for real-world usage and ensuring its
readiness for deployment. Actions include:

Reviewing and correcting any grammatical errors or inconsistencies in the

model documentation.

Highlighting key stages or actions in the deployment process by formatting

them in bold for clarity.

Deploying the model to production environments or platforms for end-user

access and utilization.

What Defines a Foundation Model?

Foundation models (FMs) stand as monumental achievements in the field of
machine learning (ML), reshaping the landscape for data scientists. Unlike
traditional approaches that require building AI systems from scratch, FMs
provide a starting point, accelerating the development of new ML applications.
Coined by researchers, the term "foundation model" refers to large neural
networks trained on extensive datasets. These models excel in a myriad of

Generative AI course 29
 tasks, including language understanding, text generation, image processing,
and natural language conversation.

The Unique Traits of Foundation Models

Foundation models possess an unparalleled adaptability, allowing them to excel
across diverse tasks with remarkable accuracy, driven by input prompts. Unlike
conventional ML models confined to specific tasks, foundation models boast a
broad spectrum of capabilities, from natural language processing (NLP) to
image classification. Their size and versatility set them apart, enabling them to
tackle complex challenges with finesse.

Applications and Use Cases

The versatility of foundation models lends itself to a multitude of applications,
including:

Enhancing customer support experiences

Facilitating language translation services

Generating diverse content

Streamlining copywriting tasks

Enabling accurate image classification

Empowering high-resolution image editing

Generative AI course 30
 Extracting insights from documents

Advancing robotics and automation

Transforming healthcare practices

Revolutionizing autonomous vehicle technologies

Challenges in Utilizing Foundation Models

Despite their prowess, foundation models present several challenges:

High infrastructure requirements and resource-intensive training processes.

Complex integration into software stacks, necessitating expertise in prompt

engineering and fine-tuning.

Limited contextual comprehension, leading to occasional inaccuracies in

responses.

Potential for unreliable or inappropriate answers, especially on unfamiliar

subjects.

Risks of bias and exposure to inappropriate content due to training data

characteristics.

Categories of Foundation Models

Foundation models encompass a variety of types, broadly classified into:

Language Models: Specialized in processing and understanding natural

language, enabling tasks like translation and text generation. Examples
include BERT, GPT-3, and T5.

Computer Vision Models: Geared towards analyzing visual data,

performing tasks such as image classification and object detection. Notable
models include ResNet, VGG, and Inception.

Multimodal Models: Designed to process both text and visual data,

facilitating tasks like image captioning and visual question answering.
Examples include DALL-E 2, Flamingo, and Florence.

Examples of Foundation Models

Generative AI course 31
 pre training :
Data Collection

General Text Data

Specialized Data

Data Preprocessing

Quality Filtering

Deduplication

Tokenization

How Does Pretraining Affect LLMs?

Mixture of Sources

Amount of Pretraining Data

Quality of Pretraining Data

Architecture for Pretraining

Encoder Decoder Architecture

Causal Decoder Architecture

Prefix Decoder Architecture

Emergent Architectures

Some Notes on Configurations

Layer Normalization

Attention

Positional Encoding

Pretraining Tasks

Language Modeling

Denoising Autoencoding

Mixture-of-Denoisers

Generative AI course 32
 Pre-training is the foundation for the power of Large Language Models (LLMs) .
By training on lots of text data, these models get good at understanding and
generating language.

Importance of Pre-training Data:

LLMs need to be trained on a lot of data to become really capable.

The size and quality of this data are super important. Good data helps
the model achieve better capabilities.

There are two main types of data used to train Large Language Models
(LLMs): general data and specialized data.

General Data:

Examples: Web pages, books, chat logs.

Why it's used: It's widely available, varied, and helps LLMs get good at
general language understanding and adaptability.

Specialized Data:

Examples: Datasets in multiple languages, scientific content, and

programming code.

Purpose: Helps LLMs become experts in specific areas or tasks.

General Text Data:

Most Large Language Models (LLMs) use general-purpose data. Let's look at
three key types:

Webpages:

What it offers: A wide range of data from the internet that provides
diverse language knowledge.

Example Dataset: CommonCrawl.

Issues: Web data can have both high-quality text (like Wikipedia) and
low-quality text (like spam). It's crucial to clean and process this data to
ensure quality.

Conversation Text:

Generative AI course 33
 Why it's useful: Helps LLMs get better at conversation and question-
answering.

Example Dataset: PushShift.io Reddit corpus.

How it's used: Conversations from online platforms are structured into
tree-like formats to capture responses. This allows multi-party chats to
be split into smaller conversations for training.

Challenges: Relying too much on dialogue data can lead to problems.

For instance, LLMs might misunderstand certain phrases as
conversation starters, affecting their response quality.

Books:

Why they matter: Books offer formal and lengthy texts which help LLMs
understand complex language structures, grasp long-term context, and
produce coherent narratives.

Example Datasets: Books3 and Bookcorpus2 found in the Pile dataset.

In essence, these general data sources help LLMs understand and
generate varied and natural language, but each source comes with its
unique strengths and challenges.

Specialized Text Data:

Specialized data helps LLMs get better at certain specific tasks. Here are three
types of specialized data:

Multilingual Text:

Purpose: Improving language understanding and generation across

multiple languages.

Example Datasets: BLOOM (covers 46 languages) and PaLM (covers

122 languages).

Benefit: These models are great at tasks like translation, multilingual

summaries, and multilingual Q&A, sometimes even outperforming
models trained just on target language data.

Scientific Text:

Why it's used: Helps LLMs understand scientific knowledge.

Generative AI course 34
 Source: Materials like arXiv papers, scientific textbooks, math websites,
and more.

Challenges & Solutions: Scientific texts have things like math symbols
and protein sequences. To handle this, they're specially tokenized and
pre-processed to fit a format LLMs can use.

Benefits: LLMs trained on scientific texts are better at scientific tasks

and reasoning.

Code:

Why it's relevant: Training LLMs on code helps with program synthesis,
a popular research area.

Current State: Even powerful LLMs, like GPT-J, find it tough to produce
good, accurate programs.

Source: Code can come from Q&A communities like Stack Exchange or
public software repositories like GitHub. This includes actual code,
comments, and documentation.

Challenges & Solutions: Code has its own syntax and logic, and is very
different from regular text. Training on code, though, might give LLMs
complex reasoning skills.

Benefits: When tasks are formatted like code, LLMs can produce more
accurate answers. In short, using specialized data gives LLMs specific
skills, from understanding multiple languages to generating code.

FIltering :

Splitting the raw text into individual tokens or subword units,

which can be fed into the model as input. This can be done using
Tokenization
various algorithms, such as whitespace-based, rule-based, or
statistical methods.

Quality Filtering Removing low-quality data from the corpus using either classifier-
based or heuristic-based approaches. Classifier-based methods
train a binary classifier with well-curated data as positive
instances and sample candidate data as negative instances, and
predict the score that measures the quality of each data example.

Generative AI course 35
 Heuristic-based methods use rules or heuristics to filter out low-
quality data based on certain criteria.

Duplicate documents can

arise from various sources,
such as web scraping, data
merging, or data
augmentation, and can lead
to several issues, such as
overfitting, bias, or
inefficiency. To address these
issues, existing studies
mostly rely on the overlap
ratio of surface features (e.g.,
words and n-grams overlap)
between documents to detect
and remove duplicate
documents containing similar
Deduplication contents. Furthermore, to
avoid the dataset
contamination problem, it is
also crucial to prevent the
overlap between the training
and evaluation sets, by
removing the possible
duplicate texts from the
training set. It has been
shown that the three levels of
deduplication (i.e.,
document-level, sentence-
level, and token-level) are
useful to improve the training
of LLMs, which should be
jointly used in practice.

How does pretraining affect LLMs?

Effect of pretraining on LLMs can be broadly categorized into three
categories

Mixture of sources

Amount of pretraining data

Generative AI course 36
 Quality of pretraining data

Pretraining Tasks
There are two commonly used pretraining tasks:

1. Language Modelling

2. Denoising Autoencoding

Language Modeling in LLMs:

Language Modeling (LM) is a foundational task in natural language
processing. It entails predicting the next token in a sequence based on its
history. When applied to Large Language Models (LLMs), especially
decoder-only models, it serves as a pivotal pre-training task

Denoising Tasks in LLM Pre-training:

The discussion revolves around denoising tasks, which are prominent pre-
training objectives for Large Language Models (LLMs). Let's break down
the content:

1. Denoising Autoencoding (DAE):

Definition: In the Denoising Autoencoding task, parts of the input

text are intentionally corrupted by replacing certain spans. The
objective is to train the model to recover the original, uncorrupted
tokens

Influence of Architecture and Pre-training Tasks on LLMs:

1. Architecture Choice:

Discussion: Early literature on pre-trained language models

extensively discussed architectural effects. However, many
LLMs use the causal decoder architecture, with limited
theoretical analysis on its advantages.

Causal Decoder and LM Objective:

Generative AI course 37
 LLMs using a causal decoder architecture with a language
modeling (LM) objective have shown strong zero-shot and
few-shot generalization capabilities.

Without multi-task fine-tuning, the causal decoder performs

better in zero-shot scenarios than other architectures.

GPT-3, a popular model, confirmed that large causal

decoders can be effective few-shot learners.

Techniques like instruction tuning and alignment tuning can

enhance the performance of large causal decoder models.

Scaling Law:

Causal decoders benefit from scaling laws: increasing

model size, dataset size, and computation can notably
improve performance.

In-depth studies on encoder-decoder models, especially at

larger scales, are needed.

Future Research:

More research is necessary to understand how architecture

and pre-training task choices affect LLM capacity. Particular
interest is in encoder-decoder architectures. Additionally,
detailed LLM configurations deserve more attention.

1. Long Context:

Context Limitation: Transformers have traditionally been

constrained by context length due to quadratic computational
costs in terms of time and memory.

Growing Demand: With increasing needs for long context

windows in tasks like PDF processing and story writing, models
are evolving. For instance, ChatGPT has expanded its context
window from 4K tokens to 16K tokens, and GPT-4 has been
extended to 32K tokens.

Extrapolation:

This refers to an LLM's ability to handle input texts that are

longer than the maximum length seen during training.

Generative AI course 38
 Position embedding techniques like RoPE and T5 bias have
displayed extrapolation abilities. For example, LMs equipped
with ALiBi have demonstrated consistent performance on
sequences much longer than training sequences.
Furthermore, the xPos method seeks to enhance the
extrapolation capability of RoPE.

Efficiency:

To address the quadratic computational challenge, various

studies have proposed more efficient attention computation
methods, such as sparse or linear attentions.

FlashAttention improves efficiency at the system level

(focusing on GPU memory IO efficiency), enabling training
LLMs with longer context windows with the same
computational budget.

Some researchers are proposing novel architectures to

tackle this efficiency challenge, such as RWKV and RetNet.

Fine-Tuning in Deep Learning

Fine-tuning is a technique in deep learning that falls under the umbrella
of transfer learning. It involves making slight adjustments to a pre-
trained model's internal parameters to adapt it to a new, related task
without starting the training process from scratch. Let's explore the
significance, steps, and common strategies associated with fine-tuning.

Significance of Fine-Tuning
Efficiency: Compared to training a model from scratch, fine-tuning
significantly reduces time and resource requirements by building
upon a pre-existing model's learned features.

Generative AI course 39
 Improved Performance: Leveraging a pre-trained model's
knowledge and representations enhances performance on task-
specific datasets, leading to better results.

Data Efficiency: Fine-tuning enables effective training even with

limited labeled data, maximizing the utility of available resources.

Steps in Fine-Tuning
1. Select a Pre-Trained Model: Choose a pre-trained model that aligns
with the nature of your task, ensuring compatibility with the required
features.

2. Adjust Model Architecture: Modify the model's architecture,

typically focusing on the top layers, to suit the specific requirements
of your task. This may involve altering output neuron counts or layer
configurations.

3. Freeze or Unfreeze Layers: Decide whether to freeze certain layers

to retain their learned representations or unfreeze them to adapt to
new data during fine-tuning, depending on the task complexity and
dataset size.

4. Training: Train the modified model on the task-specific dataset,

using a smaller learning rate compared to initial pre-training to
prevent drastic changes to learned representations while allowing
adaptation to new data.

5. Fine-Tuning Strategies: Experiment with hyperparameters, loss

functions, and training strategies to optimize fine-tuning results
iteratively.

Common Use Cases for Fine-Tuning

Fine-tuning is particularly effective in scenarios where:

Style Customization: Tailoring models to reflect specific brand

tones or qualitative aspects.

Specialization: Adapting general models to domain-specific tasks,

such as legal or medical contexts.

Few-Shot Learning: Achieving strong performance with minimal

labeled examples for specific classification tasks.

Generative AI course 40
 Addressing Edge Cases: Ensuring model robustness by handling
rare or unlikely scenarios appropriately.

Incorporating Proprietary Data: Integrating company-specific

datasets to enhance model relevance and performance.

Fine-Tuning Strategies

Full Fine-Tuning
Update the entire neural network similarly to the pre-training
process, demanding significant computational resources.

Partial Fine-Tuning
Selectively update critical parameters while freezing others to
reduce computational demands and stabilize the model.

Additive Fine-Tuning
Introduce new parameters or layers to the model and train only
these additional components while preserving the original pre-
trained weights.

Transfer Learning in Deep Learning

Transfer learning is a fundamental technique in deep learning that involves
leveraging knowledge gained from solving one task to improve
performance on a related but different task. It's a valuable strategy when
there's a scarcity of labeled data for the target task or when training
resources are limited. Let's delve into the significance, process, and various
aspects of transfer learning, akin to fine-tuning.

Significance of Transfer Learning

Efficiency: Transfer learning accelerates model training by utilizing pre-
existing knowledge, reducing the need for extensive retraining.

Improved Performance: By transferring learned features and

representations from a source task to a target task, models can achieve
better performance on new datasets.

Generative AI course 41
 Data Efficiency: Transfer learning maximizes the utility of available
labeled data by repurposing it for related tasks, making it particularly
beneficial in scenarios with limited data availability.

Process of Transfer Learning

1. Select Source Task and Model: Choose a source task and a pre-trained
model that has been trained on a relevant dataset. The model should
possess generalized features applicable to the target task.

2. Feature Extraction or Fine-Tuning: Decide between feature extraction

and fine-tuning based on the similarity between the source and target
tasks.

Feature Extraction: Extract high-level features learned by the pre-

trained model and use them as input to a new model trained for the
target task.

Fine-Tuning: Make minor adjustments to the pre-trained model's

internal parameters to adapt it to the target task, while leveraging its
learned representations.

3. Adapt Model Architecture: Modify the model's architecture as needed

to align with the requirements of the target task, such as adjusting
output neuron counts or layer configurations.

4. Transfer Learning Strategies: Experiment with different transfer

learning strategies, including frozen feature extraction, fine-tuning
specific layers, or applying domain adaptation techniques, to optimize
model performance.

Common Use Cases for Transfer Learning

Transfer learning finds application in various scenarios, including:

Domain Adaptation: Adapting models trained on one domain to perform

well in a related but different domain.

Multi-Task Learning: Training models to perform multiple tasks

simultaneously by sharing knowledge across tasks.

Zero-Shot Learning: Learning to perform tasks without any labeled

examples by leveraging knowledge from related tasks.

Generative AI course 42
 Semi-Supervised Learning: Combining labeled and unlabeled data to
improve model performance, especially in scenarios with limited labeled
data.

Transfer Learning Strategies

Frozen Feature Extraction

Extract high-level features learned by a pre-trained model and use them
as fixed inputs to a new model for the target task, preventing further
updates to the pre-trained weights.

Fine-Tuning
Adjust the internal parameters of a pre-trained model to adapt it to the
target task while leveraging its learned representations, allowing for
minor updates to the pre-trained weights.

Domain Adaptation
Incorporate techniques such as adversarial training or domain-specific
loss functions to align the learned representations of a model with the
target domain.

Differences between Transfer Learning & Fine

Tuning

Aspects Transfer Learning Full Fine-tuning

Definition In transfer learning, only a In full fine-tuning, all the

small subset of the model’s parameters of a pre-trained
parameters or a few task- model are updated during the
specific layers are trained training process. This means
while keeping the majority of the weights of every layer in the
the pre-trained model’s model are adjusted based on
parameters frozen. One of the the new training data. This

Generative AI course 43
 popular transfer learning approach can lead to significant
strategies is PEFT. changes in the model’s behavior
and performance, tailored to the
specific task or dataset at hand.

The goal is to adapt the pre-

trained model to a new task
with minimal changes to its
The goal is to comprehensively
parameters. This approach
adapt the entire pre-trained
seeks to find an optimal
model to a new task or dataset.
Objective balance between retaining the
The aim is to maximize
general knowledge acquired
performance on that specific
during pre-training and
task.
making enough task-specific
adjustments to perform well
on the new task.

Utilizes existing architecture;

With full fine-tuning, every
Model Freezes most of the layers.
parameter of the LLM gets
Architecture Only a small set of parameters
updated.
are fine-tuned.

May involve only training a

new top layer while keeping
other layers fixed. In some Involves adjusting specific
Training
cases, the number of newly layers and parameters for the
Process
trained parameters is just 1– new task.
2% of the original LLM
weights.

Data Smaller dataset with fewer Requires task-specific large

Requirement examples data set for fine-tuning.

This approach is generally

Since it involves updating all the
more resource-efficient, as
parameters, full fine-tuning is
only a small portion of the
typically more computationally
model is being updated. It
intensive and time-consuming.
requires less memory and
Computational It requires more memory and
processing power and often
Complexity processing power, as well as
leads to faster training times.
potentially longer training times,
This makes it more accessible
especially for large models. Full
for situations with limited
fine-tuning often requires a
computational resources or
large amount of GPU RAM.
for quick experimentation.

Storage Reduced storage Increased storage requirements

requirements requirements model

Generative AI course 44
 Performance can be similar,
Model Typically results in higher
but often a bit lower than full
performance performance
fine-tuning

Inference What is needed is to host

Each fine-tuned model must be
hosting original LLM and additional
hosted
Requirements model weights for inference

RAG vs Finetuning
Purpose:
Fine-tuning LLMs: Fine-tuning involves taking a pre-trained LLM (such as
GPT-3 or BERT) and adapting it to specific tasks. It is a versatile approach
used for various NLP tasks, including text classification, language
translation, sentiment analysis, and more. Fine-tuned LLMs are typically
used when the task can be accomplished with just the model itself and
doesn’t require external information retrieval.

RAG: RAG models are designed for tasks that involve both retrieval and
generation of text. They combine a retrieval mechanism, which fetches
relevant information from a large database, with a generation mechanism

Generative AI course 45
 that generates human-like text based on the retrieved information. RAG
models are often used for question-answering, document summarization,
and other tasks where accessing local information is crucial.

Architecture:
Fine-tuning LLMs: Fine-tuning LLMs usually start with a pre-trained model
(like GPT-3) and fine-tune it by training on task-specific data. The
architecture remains largely unchanged, with adjustments made to the
model’s parameters to optimize performance for the specific task.

RAG: RAG models have a hybrid architecture that combines a transformer-

based LLM (like GPT) with an external memory module that allows for
efficient retrieval from a knowledge source, such as a database or a set of
documents.

Training Data:
Fine-tuning LLMs: Fine-tuning LLMs rely on task-specific training data,
often consisting of labeled examples that match the target task, but they
don’t explicitly involve retrieval mechanisms.

RAG: RAG models are trained to handle both retrieval and generation, which
typically involves a combination of supervised data (for generation) and
data that demonstrates how to retrieve and use external information
effectively

Use Cases:
Fine-tuning LLMs: Fine-tuned LLMs are suitable for a wide range of NLP
tasks, including text classification, sentiment analysis, text generation, and
more, where the task primarily involves understanding and generating text
based on the input.

RAG: RAG models excel in scenarios where the task requires access to
external knowledge, such as open-domain question-answering, document

Generative AI course 46
 summarization, or chatbots that can provide information from a knowledge
base.

Data Collection and Preparation

Data Collection: Collect data from various sources including internet, social
media, sensors, and optimally data warehouses like Google BigQuery or
Google Cloud Storage. When collecting data to train generative AI models,
ensure that the dataset represents a diverse range of perspectives,
backgrounds, and sources. Be cautious about relying solely on data from a
single source or domain, as it may introduce bias.

Data Creation: Optionally generate/use synthetic data to augment existing

datasets to achieve better performance metrics on trained models.

Data Curation: Clean and organize data through filtering, transformation,

integration and labeling of data for supervised learning. Programmatic
labeling may be an indispensable part of how we can accelerate training
and fine-tuning.

Feature Storage: Store the gold standard of features that have been
painstakingly curated and prepped. We do this primarily for AI governance
and consistency across the enterprise and the may project that will be
attached to it; we will want to store and manage features used in training
and inference ; often distinguishing between static and dynamic or train
time and run time feature store capabilities. Looking to provide answers to
questions from development teams around: “Where did we get the data
from? Who touched the data? Did they bias the data during
curation/labeling?” For example, you can use the Vertex AI Feature Store to
support this task.

Data Bias Check. This is an essential step in the machine learning lifecycle,
especially for generative AI models, to ensure that the training data and the

Generative AI course 47
 generated content is fair, unbiased, and representative of the underlying
data distribution. Bias in generative AI models can lead to the creation of
outputs that reinforce stereotypes, discriminate against certain groups, or
misrepresent reality. We can address data bias in predictive and generative
AI, right after collecting and curating the data. By incorporating data bias
checks into the machine learning lifecycle for generative AI, we can work
towards creating models that generate fair, unbiased, and representative
content, ultimately ensuring that the AI systems we develop align with
ethical guidelines and societal values.Data Collection and Preparation

Model Training & Experimentation

Experimentation: This is name of the game in machine learning.

Pre-trained model training: Train a model using a pre-trained model as a

starting point

Fine-tune models: Update an existing model with new data or for a new
task

Model registry: This is a means and mechanism for Model Governance and
Model Risk Management. To use the Model Registry we would Store and
manage models for versioning, reuse, and auditing, marking the meta-data
of the model to see where the data came from (Data Provenance, Lineage),
how the model was trained, etc.

Distributed Training
Data parallelism: Split data across multiple devices and training them
simultaneously to speed up training

Model parallelism: Split the model across multiple devices and training them
simultaneously to fit larger models into memory

Generative AI course 48
 Federated learning: Train models collaboratively across multiple devices or
organizations while preserving privacy and security.

Model Adaptation
Transfer learning
Adapt a pre-trained model to a new domain or task

Fine-tune models
Update an existing model with new data or for a new task

AI Safety Mitigation
Ensure the model is safe, ethical, and trustworthy by mitigating misuse,
improving robustness, and ensuring privacy and security.

Distillation
Train or create a much smaller model with similar performance and
downstream task capabilities by using a student/teacher few-shot learning
approach to train a much smaller model, or, prune a model to arrive at a smaller
footprint deployable model.
There are several methods to distill large language models (LLMs) to arrive at
significantly smaller models with comparable performance. Let’s explore some
common approaches .

1. Pruning: This involves removing unimportant weights or neurons from the

LLM while preserving its accuracy. Pruning can be done using various
techniques such as magnitude-based pruning, sensitivity-based pruning,
and weight-rewinding.

2. Quantization: This involves reducing the precision of the LLM’s weights and
activations. For example, weights that were initially represented as 32-bit
floating-point numbers can be quantized to 8-bit integers. This reduces the
model’s memory footprint without significantly impacting its performance.

3. Knowledge Distillation: This involves training a smaller model to mimic the

output of the LLM on a set of training examples. The idea is to transfer the
knowledge from the larger model to the smaller one. This approach can be

Generative AI course 49
 used to train models that are orders of magnitude smaller than the original
model with only a small reduction in accuracy.

4. Low-rank Factorization: This involves decomposing the weight matrices of

the LLM into low-rank matrices. This reduces the number of parameters in
the model while preserving its accuracy.

5. Compact Embeddings: This involves reducing the dimensionality of the

input and output embeddings of the LLM. This reduces the number of
parameters in the model and speeds up inference.

6. Architectural Changes: This involves changing the architecture of the LLM

to make it more efficient. For example, the transformer architecture can be
modified to reduce its memory footprint or the number of attention heads
can be reduced.

7. Parameter Sharing: This involves sharing parameters between different

parts of the model. For example, the same weight matrix can be used for
multiple layers in a neural network, reducing the number of parameters and
improving the model’s efficiency.

Model Deployment
Let’s discuss how to serve the model using Google Cloud’s Vertex
AI endpoint. Serving a model on a Vertex AI endpoint involves packaging the
model into a container image, uploading the image to a container registry,
creating a model resource and endpoint in the Vertex AI console, deploying the
model to the endpoint, testing the endpoint, and monitoring its performance.

1. Prepare the Model. First, you need to prepare the trained model for
deployment. This typically involves packaging the model and any required
dependencies into a container image, such as a Docker image.

2. Upload the Container Image. Next, you need to upload the container image
to a container registry, such as Google Container Registry. This will allow
you to easily deploy the image to a Vertex AI endpoint.

Generative AI course 50
 3. Create a Model Resource. In the Vertex AI console, create a model
resource that will represent your deployed model. This involves specifying
the model’s name, description, and any other metadata that may be
required.

4. Create an Endpoint. Create a Vertex AI endpoint to serve your model. This

involves specifying the endpoint’s name, description, and the region where
it will be deployed.

5. Deploy the Model. Once the endpoint is created, you can deploy your
model to the endpoint. This involves specifying the container image to use
and any required settings, such as the number of replicas to deploy.

Model Maintenance
Data and model drift. Monitor and managing changes in data distribution
and model performance over time

Model versioning. Manage different versions of the model for traceability,

reproducibility, and comparison

Prompt-tuning, Prompt design: Optimize prompts used in language

generation models to improve quality, coherence, and relevance.

Performance optimization
Hardware acceleration. Optimize the model for specific hardware
accelerators to improve inference speed and efficiency

Quantization. Reduce the model size and computation requirements by

converting it to lower precision data types

Pruning. Reduce the model size and computation requirements by removing

unimportant weights and neurons.

Generative AI course 51