Deep Learning and Transformer Models

Catalog description

Deep learning—and, in particular, Transformers and large language models (LLMs)—now underpin state-of-the-art systems in vision, language, and multimodal AI, powering everything from search and coding assistants to scientific discovery. This course moves from the foundations of machine learning to modern deep learning practice: core math and optimization; MLPs and CNNs; sequence modeling before Transformers; attention and full Transformer architectures; tokenization and pretraining; scaling and adapting LLMs (LoRA, instruction tuning); alignment and post-training (RLHF/DPO/RLAIF); efficient training/inference systems (parallelism, quantization, distillation); retrieval-augmented and multimodal pipelines; rigorous evaluation and interpretability; and, finally, ethics, safety, and bias. By the end, students will (1) implement and train models from scratch and with modern frameworks, (2) design controlled experiments and evaluation suites, (3) build, adapt, and serve compact LLMs with efficiency/latency constraints, (4) apply alignment and guardrails to reduce harmful or unfaithful behavior, and (5) communicate results and ethical trade-offs through reproducible reports and model cards