Syllabus
DS 6050 · School of Data Science, University of Virginia
This course takes you from linear models to transformers and generative AI, combining rigorous mathematical foundations with from-scratch implementations in native PyTorch tensors and modules. The goal is genuine expertise rather than pattern-matching: the judgment to design, debug, and innovate in real AI systems. Lectures are mostly worked on the board; the free course book anchors modules whose chapters have completed review. The open Dive into Deep Learning textbook remains an alternative reading.
Self-paced learners: dates, submissions, and grades apply to the for-credit UVA offering (managed in Canvas). Everything you need to learn — videos, readings, notebooks, self-checks — is on this site, free.
Learning phases
Master the mathematical underpinnings and implement core algorithms from scratch with native PyTorch tensors: linear models, neural networks, and backpropagation — plus the optimization foundations and ablation methodology that recur throughout the course.
- Module 1 – Introduction to Deep Learning≈6 h · 2 h video · 2 h reading · 2 h coding
- Module 2 – Backpropagation≈6 h · 2 h video · 2 h reading · 2 h coding
- Module 3 – Optimization Foundations & Ablation Methodology≈5 h · 1.5 h video · 1.5 h reading · 2 h coding
How data modalities drive architecture: CNNs for spatial data, RNNs for sequences, attention and Transformers for universal modeling. Each module revisits the optimization challenges unique to that architecture.
- Module 4 – Convolutional Neural Networks (CNNs)≈6 h · 2 h video · 2 h reading · 2 h coding
- Module 5 – Advanced CNN Architectures≈6 h · 2 h video · 2 h reading · 2 h coding
- Module 6 – Encoder Decoder Architectures≈5 h · 1.5 h video · 1.5 h reading · 2 h coding
- Module 7 – Recurrent Neural Networks≈7 h · 2.5 h video · 2 h reading · 2.5 h coding
- Module 8 – Attention Mechanism≈6 h · 2 h video · 2 h reading · 2 h coding
- Module 9 – Transformer≈6 h · 2 h video · 2 h reading · 2 h coding
Contemporary practice: vision transformers, large-scale pretraining, parameter-efficient fine-tuning and quantization, and generative models (multimodal, diffusion, VAEs, GANs).
- Module 10 – Transformer Models in Vision and Text≈6.5 h · 2 h video · 20 min frontier outline · 2 h reading · 2 h coding
- Module 11 – Prompting, PEFT, and Quantization (Gemma)≈7 h · 2 h video · 3 h reading · 2 h coding
- Module 12 – Multimodal Learning & GenAI≈8 h · 2.5 h video · 3 h reading · 2.5 h coding
How the for-credit course is graded
Included for context; assessment is designed so the evaluation is a by-product of learning.
Programming assignments
Five assignments that progressively build implementation skill: linear models from scratch, MLPs & backpropagation, CNNs, sequence models, and attention & transformers.
Group project
A multi-phase grand challenge with formative milestones — proposal & literature review, mid-project check-in, and final deliverables with documentation and presentations.
Participation
Module quizzes (open-book, understanding over memorization) and discussion contributions, including leading one module discussion.