EECE 571F (2021 Winter Term 2): Deep Learning with Structures

Archived offering. Dates and assessment requirements below belong to this offering. View the current EECE 576L course.

Overview

Structures are pervasive in science and engineering. Some structures are conveniently observable, e.g., 3D point clouds, molecules, phylogenetic trees, social networks, whereas some are latent or hard to be measured, e.g., parse trees for languages/images, causal graphs, and latent interactions among actors in multi-agent systems. Advanced deep learning techniques have emerged recently to effectively process data in the above scenarios.

This course will teach cutting-edge deep learning models and methods with structures. In particular, for observable structures, we will introduce popular models, e.g., Transformers, Graph Neural Networks, with an emphasis on motivating applications, design principles, practical and or theoretical limitations, and future directions. For latent structures, we will introduce motivating applications, latent variable models (e.g., variational auto-encoders), and inference methods (e.g., amortization and search), and learning methods (e.g., REINFORCE and relaxation).

Announcements

• Jan. 6 2022: We will have online courses until Jan. 21.
• Jan. 12 2022: We will have online courses until Feb. 7. The plan for the weeks after will be announced later.
• Feb. 21 2022: We will have hybrid (in-class + online) courses from next week, i.e., starting from Feb. 28.

Course Information

Where and When

Pre-requisites

• Basic knowledge about machine learning, linear algebra, probability, and calculus.
• Proficiency in a programming language: preferably Python.
• Proficiency in a deep learning libraries: PyTorch, JAX, Tensorflow, etc.

Course Structure

The instructor will present the lectures every week except that students will present their projects in the last two weeks.
Students should ask all course-related questions on Piazza.

Project

Students can work on projects individually, or in groups of up to four (group should be formed as early as possible). Students are strongly encouraged to form groups via, e.g., discussing on Piazza. However, a larger group would be expected to do more than a smaller one or individuals. All students in a group will receive the same grade. Students are allowed to undertake a research project that is related to their thesis or other external projects, but the work done for this course must represent substantial additional work and cannot be submitted for grades in another course.

The grade will depend on the quality of research ideas, how well you present them in the report, how clearly you position your work relative to prior literature, how illuminating and or convincing your experiments are, and well-supported your conclusions are. Full marks will require a novel contribution.

Each group of students will write a short (>=2 pages) research project proposal, which ideally will be structured similarly to a standard paper. You don’t have to do exactly what your proposal claims - the point of the proposal is mainly to have a plan for yourself and to allow me to give you feedback. Students will do a short presentation (roughly 5 minutes for individual, 10 to 15 minutes for a larger group) for their projects towards the end of the course. At the end of the class, every group needs to submit a project report (6~8 pages).

Evaluation Policy

Grades will be based on:

• [15%] One paper reading report. Find one paper from the reading list below. Guideline & Policy, due Feb. 4
• [15%] Project proposal. Guideline & Policy, due Feb. 18
• [15%] Project presentations. Guideline & Policy
• [15%] Peer-review report of project presentations. Guideline & Policy, due Apr. 15
• [40%] Project report and code. Guideline & Policy, due Apr. 20

Important Notes

1. All reports (i.e., paper reading report, proposal, peer-review report, and final project report) must be written in NeurIPS conference format and must be submitted as PDF

2. Late work will be automatically subject to a 20% penalty and can be submitted up to 3 days after the deadline

3. UBC values academic integrity. Therefore, all students must understand the meaning and consequences of cheating, plagiarism and other academic offences under the Code of Student Conduct and Discipline.

4. It is the responsibility of each student to understand the policy for each course work, ask questions to the instructor if it is not clear, and carefully acknowledge all sources (papers, code, books, websites, individual communications) using appropriate referencing style when submitting work.

Syllabus

This is a tentative schedule, which will likely change as the course goes on.

FAQ

Can I audit or sit in?

I am very open to auditing guests if you are a member of the UBC community (registered student, staff, and/or faculty). I would appreciate that you first email me. If the in-person class is too full and running out of space, I would ask that you please allow registered students to attend.

Is there a textbook for this course?

While there is no required textbook, I recommend the following closely relevant ones for further reading:

• GNN book: “Graph Neural Networks: Foundations, Frontiers, and Applications” by Lingfei Wu, et al. Free online version
• GRL book: “Graph representation learning” by William Hamilton. Free online version
• PML book: “Probabilistic Machine Learning: An Introduction” by Kevin Murphy. Free online version
• DL book: “Deep Learning” by Ian Goodfellow, Yoshua Bengio, Aaron Courville. Free online version

I also recommend students who are self-motivated to take a look at similar courses taught at other universities:

• African Master’s in Machine Intelligence (AMMI 2021): Geometric Deep Learning
• Stanford CS224W, Winter 2021: Machine Learning with Graphs
• McGill Comp 766, Winter 2020: Graph Representation Learning
• UofT STA 4273 / CSC 2547, Spring 2018: Learning Discrete Latent Structure

Paper List

Supervised Deep Learning with Observable Structures

1. Deep sets
2. Pointnet: Deep learning on point sets for 3d classification and segmentation
3. Attention is all you need
4. An image is worth 16x16 words: Transformers for image recognition at scale.
5. Learning transferable visual models from natural language supervision
6. Sequence to sequence learning with neural networks
7. MLP-Mixer: An all-MLP Architecture for Vision
8. Semi-Supervised Classification with Graph Convolutional Networks
9. Gated Graph Sequence Neural Networks
10. How Powerful are Graph Neural Networks?
11. Spectral Networks and Locally Connected Networks on Graphs
12. NerveNet: Learning Structured Policy with Graph Neural Networks
13. The graph neural network model (the original Graph Neural Networks paper)
14. Neural Message Passing for Quantum Chemistry
15. Graph Attention Networks
16. LanczosNet: Multi-Scale Deep Graph Convolutional Networks
17. Graph Signal Processing: Overview, Challenges, and Applications
18. Convolutional Neural Networks on Graphs with Fast Localized Spectral Filtering
19. 3D Graph Neural Networks for RGBD Semantic Segmentation
20. Few-Shot Learning with Graph Neural Networks
21. Convolutional Networks on Graphs for Learning Molecular Fingerprints
22. node2vec: Scalable Feature Learning for Networks
23. Inductive Representation Learning on Large Graphs
24. Learning Lane Graph Representations for Motion Forecasting
25. Representation Learning on Graphs: Methods and Applications
26. Modeling Relational Data with Graph Convolutional Networks
27. Hierarchical Graph Representation Learning with Differentiable Pooling
28. Inference in Probabilistic Graphical Models by Graph Neural Networks
29. Do Transformers Really Perform Bad for Graph Representation?
30. Weisfeiler and Leman Go Neural: Higher-order Graph Neural Networks
31. SpAGNN: Spatially-Aware Graph Neural Networks for Relational Behavior Forecasting from Sensor Data
32. Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting
33. Geometric Deep Learning: Going beyond Euclidean data
34. Geometric Deep Learning on Graphs and Manifolds Using Mixture Model CNNs
35. Dynamic Graph CNN for Learning on Point Clouds
36. Weisfeiler and Lehman Go Cellular: CW Networks
37. Provably Powerful Graph Networks
38. Invariant and Equivariant Graph Networks
39. On Learning Sets of Symmetric Elements
40. Relational inductive biases, deep learning, and graph networks
41. Graph Matching Networks for Learning the Similarity of Graph Structured Objects
42. Deep Parametric Continuous Convolutional Neural Networks
43. Neural Execution of Graph Algorithms
44. Neural Execution Engines: Learning to Execute Subroutines
45. Learning to Represent Programs with Graphs
46. Learning to Execute Programs with Instruction Pointer Attention Graph Neural Networks
47. Pointer Graph Networks
48. Learning to Solve NP-Complete Problems - A Graph Neural Network for Decision TSP
49. Premise Selection for Theorem Proving by Deep Graph Embedding
50. Graph Representations for Higher-Order Logic and Theorem Proving
51. What Can Neural Networks Reason About?
52. Discriminative Embeddings of Latent Variable Models for Structured Data
53. Learning Combinatorial Optimization Algorithms over Graphs

Unsupervised Deep Learning with Observable Structures

1. Variational Graph Auto-Encoders
2. Deep Graph Infomax
3. GraphRNN: Generating Realistic Graphs with Deep Auto-regressive Models
4. Efficient Graph Generation with Graph Recurrent Attention Networks
5. MolGAN: An implicit generative model for small molecular graphs
6. GraphVAE: Towards Generation of Small Graphs Using Variational Autoencoders
7. Learning Deep Generative Models of Graphs
8. Permutation Invariant Graph Generation via Score-Based Generative Modeling
9. Graph Normalizing Flows
10. Constrained Graph Variational Autoencoders for Molecule Design
11. Generative Code Modeling with Graphs
12. Structured Denoising Diffusion Models in Discrete State-Spaces
13. Structured Generative Models of Natural Source Code
14. A Model to Search for Synthesizable Molecules
15. Grammar Variational Autoencoder
16. Scalable Deep Generative Modeling for Sparse Graphs
17. Energy-Based Processes for Exchangeable Data
18. Learning Discrete Energy-based Models via Auxiliary-variable Local Exploration
19. Hierarchical Generation of Molecular Graphs using Structural Motifs
20. Junction Tree Variational Autoencoder for Molecular Graph Generation

Deep Learning with Latent Structures

1. Simple statistical gradient-following algorithms for connectionist reinforcement learning (the original REINFORCE paper)
2. Neural Discrete Representation Learning
3. Categorical Reparameterization with Gumbel-Softmax
4. Neural Relational Inference for Interacting Systems
5. Contrastive Learning of Structured World Models
6. The Concrete Distribution: A Continuous Relaxation of Discrete Random Variables
7. Learning Graph Structure With A Finite-State Automaton Layer
8. Neural Turing Machines
9. Oops I Took A Gradient: Scalable Sampling for Discrete Distributions
10. Direct Policy Gradients: Direct Optimization of Policies in Discrete Action Spaces
11. Gradient Estimation with Stochastic Softmax Tricks
12. Differentiation of Blackbox Combinatorial Solvers
13. REBAR: Low-variance, unbiased gradient estimates for discrete latent variable models
14. Monte Carlo Gradient Estimation in Machine Learning
15. Backpropagation through the Void: Optimizing control variates for black-box gradient estimation
16. Thinking Fast and Slow with Deep Learning and Tree Search
17. Mastering the Game of Go without Human Knowledge
18. Memory-Augmented Monte Carlo Tree Search
19. M-Walk: Learning to Walk over Graphs using Monte Carlo Tree Search
20. DSDNet: Deep Structured Self-driving Network
21. Learning to Search with MCTSnets
22. Direct Loss Minimization for Structured Prediction
23. Stochastic Beams and Where to Find Them: The Gumbel-Top-k Trick for Sampling Sequences Without Replacement
24. Direct Optimization through argmax for Discrete Variational Auto-Encoder
25. Learning Compositional Neural Programs with Recursive Tree Search and Planning
26. Reinforcement Learning Neural Turing Machines - Revised
27. The Generalized Reparameterization Gradient
28. Gradient Estimation Using Stochastic Computation Graphs
29. Learning to Search Better than Your Teacher
30. Learning to Search in Branch-and-Bound Algorithms
31. Model-Based Planning with Discrete and Continuous Actions
32. Learning Transferable Graph Exploration
33. Retro*: Learning Retrosynthetic Planning with Neural Guided A* Search