Generative AI and Large Language Models

Explore how generative ai, a subset of deep learning, generates novel content—text, images, video, or audio—by learning input–output relationships, unlike conventional ai models that only make predictions.

Learn the difference between generative and discriminative models, including how discriminative models classify inputs and how generative models generate new, similar inputs by sampling distributions.

Review basic probability terms such as conditional, joint, and marginal probabilities, and connect them to generative and discriminative models via Bayes theorem.

Map probability concepts to discriminative and generative models through digit recognition, showing how handwritten digits are classified and how generative models produce similar images with confidence.

Compare discriminative and generative models by modeling conditional probability versus learning the joint distribution p(x, y) and sampling from it for both classification and generation.

Contrast discriminative and generative models by contrasting conditional probability with joint probability, showing how generative models sample from p(x,y) to create data, while discriminative models define decision boundaries for classification.

Explore how generative AI differs from discriminative models and learn how large language models generate and understand human-like text, trained on vast internet data with billions of parameters.

Explore why language understanding is hard due to lexical, syntactic, referential, and semantic ambiguities, and how large language models interpret and resolve these challenges.

Explore how a large language model uses a context window and prompts to generate responses, with a quick demo of ChatGPT and API-based prompt processing.

Understand how a large language model predicts the next token from a probability distribution, using sampling strategies from maximum probability to creative randomness, guided by encoder-decoder architecture.

Explore applications of generative ai and large language models, including content generation, translation, text summarization, code generation and completion, and building a basic chatbot for tools and applications.

Explore how fully connected networks learn: weights store training in interconnections, layers process data, and gradient descent minimizes error across input, hidden, and output layers with activation for nonlinearity.

Discover how artificial neural networks function as data-driven function approximators that process input, infer patterns, and predict outputs like house prices by learning from training data.

Explore why a simple fully connected neural network struggles with sequence data and how recurrent neural networks address context and dependencies, leading to transformers and large language models.

Learn how recurrent neural networks use a feedback loop to remember past outputs, handling sequential data; the present state h_t links current input x_t with prior outputs.

Compare pretraining from scratch with finetuning a pre-trained model for domain adaptation, and learn how foundational models, in-context learning, and weight matrices shape practical compute and memory needs.

Explore how transformers convert text into vectors via tokenization into tokens, word and positional embeddings, an encoder–decoder architecture, and a softmax output, using GPT-2 vocabulary.

Explore tokenization using the transformers library and a GPT-2 pre-trained model, mapping text to model tokens and token IDs with auto tokenizer.

Learn how tokens become embeddings by mapping words to vectors, capturing similarity with dot products. Explore positional embedding to preserve word order in a high-dimensional space.

Understand how tokenization, embedding, and positional encoding feed an encoder–decoder transformer with attention to generate contextually next-word predictions.

Examine self-attention in transformers and the QKV matrix, detailing query, key, and value vectors, attention scores via dot products and softmax, and how these weights drive multi-headed encoder blocks.

Demonstrates how to generate embedding vectors from aGPT-2 model using the GPT-2 tokenizer and auto model, and to extract the last hidden state embeddings in Python.

Build a chatbot from scratch using a pre-trained dialogue GPT medium model from Microsoft via Hugging Face Transformers; tokenize, encode, generate, and test with a Python lab demo.

Explore how temperature, top k, and top p control randomness and diversity in large language model outputs, highlighting their interdependence and impact on precision versus creativity.

Explore how adjusting inference parameters such as temperature and top tokens affects model creativity and output, and learn to run different settings to produce varied or narrow responses.

Learn sentiment analysis with a distilbert base uncased model via the transformers pipeline to classify text as positive, negative, or neutral, with context aware scoring.

Build a simple English-to-French translator using transformers and a pipeline, selecting models from Hugging Face and tuning inference parameters.

Learn how the rouge score evaluates gisting evaluation by comparing candidate and reference tokens, using recall-oriented, unigram and n-gram clipping, with practical Python examples.

Explore in-context learning in large language models, adapt prompts via prompt engineering, and compare zero shot, one shot, and few shot inference to elicit better responses without retraining.