This repository contains a collection of projects from the course "Deep Learning and LLM based Generative Al Systems" (CSCI-GA 3033-091). The projects demonstrate a range of skills in deep learning, from fundamental concepts to advanced applications of Large Language Models (LLMs).
This course provides a graduate-level introduction to Deep Learning systems, with an emphasis on practical system performance issues and related research. The course covers several topics related to Deep Learning (DL) systems and their performance. Both algorithmic and system-related building blocks of DL systems are covered, including DL training algorithms, network architectures, and best practices for performance optimization. The latter half of the course has an in-depth exploration of Large Language Models (LLMs), covering key areas towards advanced topics including attention mechanisms, transformer models, prompt engineering, LLM applications, pre-training strategies, Reinforcement Learning with Human Feedback (RLHF), efficient LLM serving techniques, fine-tuning methods, and benchmarking specifically for LLMs.
This repository is organized into the following projects:
-
Project 1: Bias-Variance Tradeoff, Regularization, and Learning Rate Policies: An exploration of fundamental machine learning concepts, including the bias-variance tradeoff, regularization techniques like L2 regularization, and the impact of different learning rate policies.
-
Project 2: CNNs and I/O Optimization: A project focused on Convolutional Neural Networks (CNNs) and the optimization of data input/output (I/O) for efficient training. This project also delves into the use of classifiers like AdaBoost and Logistic Regression.
-
Project 3: Single Shot MultiBox Detector (SSD), ONNX Model, and ORT Inferencing: This project explores the Single Shot MultiBox Detector (SSD) model for object detection. It covers the process of fine-tuning a pre-trained model, converting it to the ONNX format, and deploying it for inference using the ONNX Runtime (ORT).
-
Project 4: Approximate Nearest Neighbors (ANN) using Hierarchical Navigable Small World (HNSW): An implementation and analysis of the Hierarchical Navigable Small World (HNSW) algorithm for Approximate Nearest Neighbor (ANN) search, a key technique for efficient vector search in large datasets.
-
Project 5: LLM Summarization, Benchmarking, and Fine-Tuning: A three-part project covering the evaluation of LLM-generated text using the ROUGE-L score, benchmarking LLM inference performance with vLLM, and fine-tuning a generative AI model for dialogue summarization using techniques like Parameter-Efficient Fine-Tuning (PEFT) and Low-Rank Adaptation (LoRA).