Guideline & Policy for Course Projects
Students must work on projects individually. 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
The grade will depend on the quality of the ideas, how well you present them in the report, how illuminating and or convincing your experiments are, and well-supported your conclusions are.
The project report should be a manageable amount of work, e.g., reproducing an existing model or surveying a research topic.
Length (5%)
It should be 4 to 8 pages, not including appendices or bibliography.
Don’t be afraid to keep the text short and to the point, and to include large illustrative figures.
Code & Appendix (15%)
You should submit the compressed file (e.g., in zip format) of PDF and code unless you are doing a pure theoretic project.
If that is the case, you should make sure you submit the appendix that include all the proof.
You can include as many proofs, extra details, experiments, etc. as you want in the appendices.
Abstract (10%)
It should summarize the main idea of the project and its contributions.
Introduction (15%)
It should clearly state the problem being addressed and or the method being reproduced and why it is important.
Model/Method (30%)
The idea is to make your paper more accessible, especially to readers who are starting by skimming your paper.
Here are a few important tips:
- It is very important to include a figure to illustrate the main computation graph of the model.
A nice figure would get you some bonus.
You must create a new figure, not just use someone else’s, even with attribution!
- Equations are very helpful if you use notations rigorously and concisely.
- Algorithm box is also very useful when your proposed method is hard to parse from pure texts.
- Formal description of the models, loss functions, conjectures, problem domains, theorems, propositions, etc.
- Highlight how your model is different from other approaches via, e.g., using figures or tables.
Experiments (20%)
It should contain one or more from the following list:
- A comparison of your reimplemented model/method with the original and other baselines on at least one real-world datasets.
- An ablation study on specific design choices.
- Detailed descriptions of datasets (e.g., how they are collected, key statistics, and properties), evaluation metrics, how you trained your model, and any tricks you used to get it to work.
- Quantitative and or qualitative analysis of experimental results.
- If doing a survey, include a table comparing the properties of the different approaches.
Conclusion & Future Work (5%)
It should consist of the main takeaways of your project.
It should also include a discussion on the limitations and potential future directions to improve.
Recommend Paper List
- Deep sets
- Pointnet: Deep learning on point sets for 3d classification and segmentation
- Attention is all you need
- An image is worth 16x16 words: Transformers for image recognition at scale.
- Learning transferable visual models from natural language supervision
- Sequence to sequence learning with neural networks
- MLP-Mixer: An all-MLP Architecture for Vision
- Semi-Supervised Classification with Graph Convolutional Networks
- Gated Graph Sequence Neural Networks
- How Powerful are Graph Neural Networks?
- Spectral Networks and Locally Connected Networks on Graphs
- NerveNet: Learning Structured Policy with Graph Neural Networks
- The graph neural network model (the original Graph Neural Networks paper)
- Neural Message Passing for Quantum Chemistry
- Graph Attention Networks
- LanczosNet: Multi-Scale Deep Graph Convolutional Networks
- Graph Signal Processing: Overview, Challenges, and Applications
- Convolutional Neural Networks on Graphs with Fast Localized Spectral Filtering
- 3D Graph Neural Networks for RGBD Semantic Segmentation
- Few-Shot Learning with Graph Neural Networks
- Convolutional Networks on Graphs for Learning Molecular Fingerprints
- node2vec: Scalable Feature Learning for Networks
- Inductive Representation Learning on Large Graphs
- Learning Lane Graph Representations for Motion Forecasting
- Representation Learning on Graphs: Methods and Applications
- Modeling Relational Data with Graph Convolutional Networks
- Hierarchical Graph Representation Learning with Differentiable Pooling
- Inference in Probabilistic Graphical Models by Graph Neural Networks
- Do Transformers Really Perform Bad for Graph Representation?
- Weisfeiler and Leman Go Neural: Higher-order Graph Neural Networks
- SpAGNN: Spatially-Aware Graph Neural Networks for Relational Behavior Forecasting from Sensor Data
- Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting
- Geometric Deep Learning: Going beyond Euclidean data
- Geometric Deep Learning on Graphs and Manifolds Using Mixture Model CNNs
- Dynamic Graph CNN for Learning on Point Clouds
- Weisfeiler and Lehman Go Cellular: CW Networks
- Provably Powerful Graph Networks
- Invariant and Equivariant Graph Networks
- On Learning Sets of Symmetric Elements
- Relational inductive biases, deep learning, and graph networks
- Graph Matching Networks for Learning the Similarity of Graph Structured Objects
- Deep Parametric Continuous Convolutional Neural Networks
- Neural Execution of Graph Algorithms
- Neural Execution Engines: Learning to Execute Subroutines
- Learning to Represent Programs with Graphs
- Learning to Execute Programs with Instruction Pointer Attention Graph Neural Networks
- Pointer Graph Networks
- Learning to Solve NP-Complete Problems - A Graph Neural Network for Decision TSP
- Premise Selection for Theorem Proving by Deep Graph Embedding
- Graph Representations for Higher-Order Logic and Theorem Proving
- What Can Neural Networks Reason About?
- Discriminative Embeddings of Latent Variable Models for Structured Data
- Learning Combinatorial Optimization Algorithms over Graphs
- On Layer Normalization in the Transformer Architecture
- Swin Transformer: Hierarchical Vision Transformer using Shifted Windows
- Recipe for a General, Powerful, Scalable Graph Transformer
- Variational Graph Auto-Encoders
- Deep Graph Infomax
- GraphRNN: Generating Realistic Graphs with Deep Auto-regressive Models
- Efficient Graph Generation with Graph Recurrent Attention Networks
- MolGAN: An implicit generative model for small molecular graphs
- GraphVAE: Towards Generation of Small Graphs Using Variational Autoencoders
- Learning Deep Generative Models of Graphs
- Permutation Invariant Graph Generation via Score-Based Generative Modeling
- Graph Normalizing Flows
- Constrained Graph Variational Autoencoders for Molecule Design
- Generative Code Modeling with Graphs
- Structured Denoising Diffusion Models in Discrete State-Spaces
- Structured Generative Models of Natural Source Code
- A Model to Search for Synthesizable Molecules
- Grammar Variational Autoencoder
- Scalable Deep Generative Modeling for Sparse Graphs
- Energy-Based Processes for Exchangeable Data
- Learning Discrete Energy-based Models via Auxiliary-variable Local Exploration
- Hierarchical Generation of Molecular Graphs using Structural Motifs
- Junction Tree Variational Autoencoder for Molecular Graph Generation
- Simple statistical gradient-following algorithms for connectionist reinforcement learning (the original REINFORCE paper)
- Neural Discrete Representation Learning
- Categorical Reparameterization with Gumbel-Softmax
- Neural Relational Inference for Interacting Systems
- Contrastive Learning of Structured World Models
- The Concrete Distribution: A Continuous Relaxation of Discrete Random Variables
- Learning Graph Structure With A Finite-State Automaton Layer
- Neural Turing Machines
- Oops I Took A Gradient: Scalable Sampling for Discrete Distributions
- Direct Policy Gradients: Direct Optimization of Policies in Discrete Action Spaces
- Gradient Estimation with Stochastic Softmax Tricks
- Differentiation of Blackbox Combinatorial Solvers
- REBAR: Low-variance, unbiased gradient estimates for discrete latent variable models
- DSDNet: Deep Structured Self-driving Network
- Learning to Search with MCTSnets
- Direct Loss Minimization for Structured Prediction
- Stochastic Beams and Where to Find Them: The Gumbel-Top-k Trick for Sampling Sequences Without Replacement
- Direct Optimization through argmax for Discrete Variational Auto-Encoder
- Learning Compositional Neural Programs with Recursive Tree Search and Planning
- Reinforcement Learning Neural Turing Machines - Revised
- The Generalized Reparameterization Gradient
- Gradient Estimation Using Stochastic Computation Graphs
- Learning to Search Better than Your Teacher
- Learning to Search in Branch-and-Bound Algorithms
- Model-Based Planning with Discrete and Continuous Actions
- Learning Transferable Graph Exploration
- Retro*: Learning Retrosynthetic Planning with Neural Guided A* Search
- Monte Carlo Gradient Estimation in Machine Learning
- Backpropagation through the Void: Optimizing control variates for black-box gradient estimation
- Thinking Fast and Slow with Deep Learning and Tree Search
- Mastering the Game of Go without Human Knowledge
- Memory-Augmented Monte Carlo Tree Search
- M-Walk: Learning to Walk over Graphs using Monte Carlo Tree Search