Course Details
Formerly offered as EECE 571F. From Fall 2026 onwards, the course is offered as EECE 576L: Advanced Topics in Deep Learning.
Detailed Overview
EECE 576L is a graduate course about the principles that shape modern deep learning systems. It connects three areas—geometric deep learning, probabilistic and generative modeling, and deep reinforcement learning—and emphasizes the assumptions and objectives that connect these model families.
The geometric deep learning portion begins with invariance and equivariance and develops neural architectures for sets, sequences, graphs, and data with group symmetries. We study set models, attention and Transformers, message-passing and graph-convolutional networks, and group-equivariant and steerable networks. The goal is to understand how data structure and symmetry guide parameter sharing and model design.
The probabilistic and generative modeling portion connects autoregressive language models with diffusion, score-based, and flow models. We compare their training objectives, parameterizations, and generation procedures, building a common view of several major approaches to modern generative modeling.
The reinforcement learning portion introduces policy-gradient methods and learning from feedback for language-model post-training. Student presentations examine methods such as GRPO, DPO, KTO, and learning from verifiable rewards, providing concrete examples of how optimization objectives shape learning dynamics.
The course combines instructor lectures, student-led presentations of recent research papers, and a substantial research project. Students practice identifying assumptions, evaluating evidence, explaining contributions and limitations, selecting appropriate baselines, and supporting conclusions with reproducible experiments or rigorous theory.
Learning Outcomes
By the end of the course, students should be able to:
- Explain and compare the inductive biases of set, sequence, graph, and group-equivariant models.
- Derive and interpret core objectives for autoregressive, diffusion, flow, and policy-gradient methods.
- Critically evaluate research papers and communicate their assumptions, contributions, and limitations.
- Design and carry out a research project with appropriate baselines, reproducible experiments or rigorous theory, and evidence-supported conclusions.
Prerequisites
- Basic knowledge of machine learning, linear algebra, probability, and calculus.
- Proficiency in Python or another programming language.
- Proficiency in a deep learning library such as PyTorch or JAX.
Grading
- 30%: Paper presentation
- 10%: Project proposal, deadline TBA
- 10%: Project presentation
- 10%: Presentation peer review, deadline TBA
- 40%: Project report and code, deadline TBA
- Up to 3 percentage points: Participation
See the course project overview for group, scope, and deliverable expectations and the course policies for rules that apply to assessments.