Overview
The course has an open-ended project, with the objective of giving you an
opportunity to learn something hands-on. My expectation on your projects is
high. It is not because that I want to give you a hard time, but it is
because that I hope to bring you one step closer to the machine learning
community. (Yes, by the machine learning community I really mean the real
one. There are many gobbly goop in our times, and you know what I mean.) If your career goal is to pursue a data science, computer vision, or
deep learning type of jobs, it is better to understand their expectations sooner
than later.
With this goal in mind, I will aim for a standard in par with ICML, NeurIPS,
ICLR, CVPR, ICCV, ECCV, etc. I will ask you to write a 10-page paper. The
teaching staff (I and the TAs) will act like the reviewers of your paper.
You need to choose one of the topics below. There are specific suggestions for
each topic.
Instructions
Your report (aka paper) should have the following sections.
Introduction. In the introduction section, you need to define the
problem, and justify why it is a valuable problem to investigate. In each of
the topics below, you are likely going to pick one or two sub-topics to
investigate. In the introduction, I want to see your explanation of why do
you pick those sub-problems. You can explain their significance, e.g., by
solving this problem we will gain certain insights. Think about a conference
reviewer. Why would somebody accept your paper? It has to be relevant, and
useful.
Related work. This is the literature review section. Demonstrate to me
that you have read several papers, and you are able to summarize them into
meaningful categories. A good literature survey should articulate the
limitations of the existing work, and highlight the new findings of your
work. Try not to give a laundry list of papers, because they are not very
useful.
Method. The heading of this section is up to you. I call it “method”, but
you can call it whatever you want. This is the main part of your paper. If
you are proposing a new idea, you need to explain your idea. If you are
studying some phenonomenon, you need to explain the insights behind the
phenonomenon. A good method section should contain a few very carefully drawn
figures to illustrate your ideas. Depending on the nature of your project,
you may want to combine this section with the experiment section. If this
appears necessary for your work, please make your best judgement.
Experiment. As the heading suggests, this is the place where you put all
your experimental results. If you are comparing with other methods, you need
to define the evaluation metric. If you conduct an ablation study, explain
why your ablation study is fair and meaningful. Memember, experiments are
presented in order to show evidence of anything you claim. If you claim
something, you'd better have an experiment to support.
Conclusion. People are sometimes confused about the conclusion section.
In my opinion, conclusion is not the same as summary. Yes, you need to
summarize the paper in this section. But more importantly, you want to
explain the limitations of your findings. You also need to suggest future
directions of your work. Sometimes, you may have found some unexpected
outcomes. Then in the conclusion you may want to comment on them.
Every paper is unique, and so you need to make the best judgement of what to
include and what not to include in the paper. The above outline is
recommended, but not mandatory. However, based on my 15+ years experience in
reviewing papers, serving as program chairs and journal editors, the
recommended outline is quite robust. So if it is your first time writing,
please consider it.
Deadlines
The project will be graded in a two-phase process.
Phase 1: Initial Submission (Optional)
Around Fall break. Check Brightspace for the exact deadline.
This submission is voluntary.
If you like to receive initial feedback from the TAs, you can submit in this phase.
The TAs will give you some rough suggestions, of whether you are on track, or if there are things you need to improve.
To participate in the initial submission cycle, you need to really write a full paper. If you submit only an abstract or a half-way done paper, we will not offer any feedback.
The TAs will read and offer suggestions.
Please submit through gradescope.
Phase 2: Final Submission
Last day of the semester. Check Brightspace for the exact deadline.
Hard deadline. No extension.
This is the paper that we will grade.
Please submit through gradescope.
Grading Criteria
Review Process:
Each report will be graded by 3 graders (all my senior PhD students), independently.
Each TA will give you a score: Excellent (9-10), Good (7-8), Borderline (5-6), Weak (3-4), Bad (0-2).
We will take the average
Specific Criteria:
How well do you understand the literature?
What are the new findings you have?
Is there any innovative ideas?
How deep is your analysis?
Are the experimental results complete?
Do you have justification for things you claim?
Is your paper easy to read? Are your ideas elaborated clearly?
Research opportunities beyond ECE 50024:
Depending on how much you have accomplished and how much innovations you have demonstrated, I may encourage you to submit your project paper to real conferences.
I am always looking for good students. This could be an opportunity for you to demonstrate your capability.
Submission Format
Please use the official ICML 2023 LaTeX template to type your report.
You can download the template at
https://icml.cc/Conferences/2023/StyleAuthorInstructions
Page length: no more than 10 pages. References do not count towards the 10 pages.
No supplementary materials. All reports have to be self contained.
Do not change the margin, font size, etc.
This is an individual project. If you
work with a friend, acknowledge the person. You need to write your own
report.
List of Recommended Papers
Deep Learning
Denoising Diffusion Probabilistic Models [pdf]: Training a network to reduce noise iteratively yields a high-quality generative model
Neural Ordinary Differential Equations [pdf]: By interpreting the layers of the neural network as the differential of a function, the number of deep network layers becomes data dependent. This effectively yields a “continuous depth” network, and the major contribution of the paper is to make this computation feasible.
Demystifying MMD GANs [pdf]: This paper shows gradient estimates for MMD GANs and Wasserstein GANs are unbiased, but learning a discriminator based on samples leads to biased gradients for the generator parameters. This paper is an accessible machine learning theory paper. See also: Maximum Mean Discrepancy [pdf], Peason Chi2 [Wiki], & Related Talks [Video, Slides]
Variational Inference with Normalizing Flows [pdf]: This paper enables a class of deep learning networks (named normalizing flows) to be the approximate distribution used in variational inference. See also: Amortized Inference Regularization [pdf].
Deep State Space Models for Time Series Forecasting [pdf]: A novel approach to probabilistic time series forecasting that combines state space models with deep learning.
Noise2Score [pdf]: This paper presents a method for the unsupervised training of deep denoiser networks using only noisy images.
Graph Contrastive Learning with Augmentations [pdf]: This paper proposes a graph contrastive learning (GraphCL) framework for learning unsupervised representations of graph data.
Classical Topics
Plug-and-Play ADMM for Image Restoration: Fixed Point Convergence and Applications [pdf]: This paper presents a Plug-and-Play ADMM algorithm with provable fixed point convergence for image restoration problems.
MCMC for doubly-intractable distributions [pdf]: This paper presents an MCMC algorithm for sampling the posterior distribution when the likelihood is intractable. See also their referenced SAVM method pdf.
Bayesian Learning via Stochastic Gradient Langevin Dynamics [pdf]: This paper proposes a framework for learning from large-scale datasets where iterates converge to samples from the true posterior distribution.
Applications
DeepFaceLab: Integrated, flexible, and extensible face-swapping framework [pdf]: To build deep fake detection methods, one must understand existing tools to create deep fakes. This project, named DeepFaceLab, provides a framework for face-swapping.
GODEL: Large-Scale Pre-training for Goal-Directed Dialog [pdf]: This project is an open-source high-quality large language model from Microsoft.
AudioLDM: Text-to-Audio Generation with Latent Diffusion Models [pdf]: This paper proposes a Text-to-Audio (TTA) system that is built on a latent space to learn the continuous audio representations from contrastive language-audio pretraining (CLAP) latents.
Topics
You can choose from the list of recommended papers above
Or, you can pick your own paper. Criteria when choosing a paper:
|