Generative Artificial Intelligence K Aditya
• Overview of Gen AI
• What are LLMs
• Different ways of interacting with
LLMs
• Prompting
• Prompting techniques
• Transfer learning
• Fine tuning
• Retrieval Argumental generation
Brief Overview of AI to Gen AI
evolution
Era Timeframe Description
The concept of artificial intelligence is conceived, with early discussions and theories on how machines could simulate aspects of
Early Beginnings 1950s
human intelligence.
Formalization of AI 1960s Formal academic fields are established; foundational algorithms like the perceptron are developed. Focus on logic and reasoning.
Rule-based systems are developed to emulate the decision-making ability of a human expert, applied in areas such as medical
Expert Systems 1970s
diagnosis.
Focus shifts to statistical methods for data analysis; algorithms such as decision trees and early neural networks are used to
Machine Learning Era 1980s
recognize data patterns.
Advancements in neural network research, hindered in part by hardware limitations but setting the stage for future deep learning
Rise of Neural Networks 1990s
approaches.
Support Vector Machines 2000s The introduction and popularization of support vector machines and other kernel methods revolutionize pattern recognition.
Artificial neural networks, particularly deep learning architectures, take center stage, driven by increased data availability and
Deep Learning Revolution 2010s
powerful GPUs.
Ian Goodfellow introduces Generative Adversarial Networks (GANs), pushing forward the generation of realistic images, videos, and
Generative AI 2016
audio.
Introduction of the Transformer architecture, leading to the development of OpenAI’s first Generative Pre-trained Transformer (GPT-
Transformers & GPT-1 2017
1), enhancing natural language understanding and generation.
OpenAI releases GPT-2, notable for its ability to generate coherent text passages and for initially being deemed too dangerous to
GPT-2 2019
release fully due to potential misuse.
Launch of GPT-3 by OpenAI, featuring 175 billion parameters, setting a new standard for language models in versatility and
GPT-3 2020
capability.
Overview to Generative AI
• Generative AI refers to artificial intelligence models that can generate
new data instances that resemble the training data, including text,
images, audio, and video
• Primarily utilizes deep learning architectures such as Generative
Adversarial Networks (GANs), Variational Autoencoders (VAEs), and
Transformer-based models like GPT (Generative Pre-trained
Transformer).
• Employs advanced techniques in unsupervised, semi-supervised, and
supervised learning to train models on large datasets, enabling them
to produce high-quality, novel outputs.
Discriminative Vs Generative model
Discriminative Models Generative Models
• Machine Learning models: • Generative Models:
• Logistic Regression, Decision Trees, • Generative Adversarial Networks (GANs),
Support Vector Machines Variational Autoencoders (VAEs),
Transformer Models (e.g., GPT-3)
• Used for regression, classification,
ranking problem statements • Used for image generation, text
generation, synthetic data augmentation
• Artificial Neural Network (ANN): • Sequence to Sequence Models:
• Multilayer Perceptron (MLP), Recurrent
Neural Network (RNN), Radial Basis • Basic Seq2Seq, LSTM-based Seq2Seq,
Function (RBF) Network Transformer-based Seq2Seq
• Used for image classification, text • Used for machine translation, speech
classification, time series prediction recognition, text summarization
• Convolutional Neural Networks:
• LeNet, AlexNet, VGG-Net
• Used for image classification, object
detection, video analysis
Large Language Models
• Advanced AI Models: LLMs are designed to mimic human-like
understanding of language.
• Deep Learning Techniques: Utilize sophisticated deep learning methods to
analyze and generate language.
• Transformer Architectures: Employ transformers that use self-attention
mechanisms to process text, understanding the context and relevance of
each word.
• Learning from Data: Trained on vast amounts of diverse text data to learn
linguistic patterns, structures, and nuances.
• Text Generation: Capable of generating coherent and contextually
appropriate text by predicting the next word in sequences.
• Fine-Tuning: Can be specifically fine-tuned on smaller, task-oriented
datasets to optimize performance for particular applications.
• Diverse Applications: Used in creating chatbots, writing assistants, content
generation tools, and for tasks such as summarization and translation.
Interacting with LLMs
• Interaction through Apps : User friendly apps
• GPT4ALL
• AI Navigator (Anaconda)
• Pieces
• Web interfaces : Simple and accessible on all different platforms
• Chatgpt
• Anthropic
• Perplexity
• APIs : Integrated with larger app. Gives more control
• Cloud solution (AWS, Google AI studio, Azure)
• OpenAI
Prompting and Prompting Engineering
• Prompts:
• It’s a means to interact with the LLMs
• Set of instructions provided to a LLM for a to model respond
• Context
• Tone
• Style ..
• Better the prompts the betters the response
• Prompt Engineering:
• Its about the right question in right way
• A rule book for creating better prompts
Different Prompting Techniques
• Fine tuning Pattern
• Cognitive verifier pattern
• Question refinement pattern
• Game play pattern
Fine Tuning and Transfer learning
• Transfer learning
• The process of taking a pretrained model to train just the last layers to fit to
our requirements
• Very recourse intensive
• Lots of data needed
• Finetuning
• The process of improving the model with minimal weight change for the given
problem
• Less data is sufficient
• Less computationally intensive
RAGs (Retrieval-Augmented Generation )
• Retrieval-Augmented Generation (RAG): Combines retrieval of
relevant documents or knowledge snippets with generative language
models to produce contextually accurate and detailed responses.
• Enhanced Accuracy: RAG retrieves specific information from a
knowledge base, improving response accuracy compared to
generation from general data alone.
• Applications: Used in Q&A systems, chatbots, and knowledge
management systems where accurate, context-driven information is
required.
Thank You