Complete your 1-page project proposal on a separate sheet (in LaTeX or another word processor) and submit as a group on Gradescope. Late days can be used for this project proposal, but not submitting this checkpoint will result in a grade of 0 for the final project. So please submit something on time!
Note: The project should be done in a group of 2 or 3 students from the class, with exceptions for 1-person projects only if you have explicitly cleared it with the instructor first. If you have not been able to find a project group before the deadline, you must submit an individual proposal; if you have not been approved for an individual project, the course staff will match you with other students based on your proposal content.
Project types
The project can take any of the following forms:
High quality software implementation of algorithms related to the class that are currently not publicly available, to be released for public usage.
Application of a machine learning model to a previously unconsidered dataset and to a specific scientific question. For this: find some interesting (for you) data. Keep in mind the issue of stationarity. Make sure you have enough data to be able to fit a reasonable model (the more parameters, the more data you’ll need).
Extension to an existing method, or theoretical analysis of an existing algorithm. This would likely have a scope outside the course, e.g., it could be the starting point for a longer research project.
Multiple models comparison on an existing dataset.
If you are unsure if your idea fits into these types, talk to the instructor.
Proposal format
Write a proposal that details the question you are planning to address, which dataset are you planning to use (if applicable), the family of algorithms used for the analysis, and how you plan to evaluate these methods. The goal is to check that you have a plan, so add whatever details you have already worked out that may be relevant (within the space limits: no more than 1 page).
Important: Explain also how the tasks will be allocated across team members (who does what).
Project ideas
If anything is unclear, or you want to discuss your idea, please come to the instructor’s office hours to discuss.
Past project examples
Model comparison
- HMMs vs RNNs for anomaly detection
- Automatic speech recognition - HMMs Acoustic models, Deep nets
Implementation
- RNNs for multivariate time series with missing values (GLU-R)
- Deep Kalman filtering
- Object tracking: implementation + model comparison
Dataset applications
- Modeling protein sequences + Model comparison HMMs vs RNNs
- Forecasting Wikipedia web traffic
- Prediction of Beijing pollution levels
- NYC environmental sounds (check out J. Bello’s page)
- EKG based prediction
- Emotion detection in audio
- Cryptocurrency modeling
- Hourly electrical demand
- Earthquake prediction
New algorithms/theory
- Gaussian Process Factor Analysis with Spectral Mixture kernels
- The role of dependent measurements in causal inference
Creative applications
- Voice synthesis: speak like Pippa the pig
- Image synthesis: video style transfer
New ideas to explore
The following are some new ideas to explore, but feel free to come up with your own! The instructor will keep adding to this list as new ideas come up.
LLM time series forecasting. Investigate if Large Language Models can perform zero-shot time series forecasting by treating numbers as text (Gruver et al., 2024).
Probabilistic models for neural data. Using data from Neuromatch, apply probabilistic models to synthesize and analyze realistic neural spike trains.
Neural network training trajectories. Analyze the geometric properties of a neural network’s training trajectory through its loss landscape to better understand optimization dynamics (Hu et al., 2023).