VS265 project ideas

Project ideas

I'm an undergrad in Physics interested in theoretical neuroscience and machine learning.

Some project ideas:

-Biology inspired music compression with Minimum Probability Flow learning and hopfield networks. 
Image compression has been achieved with these techniques; adapting the 
technique to music compression by mirroring the signal processing of the cochlea would be interesting. 
Teammates with strong EECS backgrounds would be desired for this project. 

I'd also be interested in using Minimum Probability Flow learning in other ways. It's a really efficient
learning rule that would make the training of really large networks tractable. 
Minimum Probability Flow Learning: http://arxiv.org/abs/0906.4779
Robust exponential storage in Hopfield networks: http://arxiv.org/abs/1206.2081


- Hamiltonian Monte Carlo sampling in continuous time. 
This would be an extension of a recent paper on an efficient way to sample without 
detailed balance authored by Mayur and Jascha Sohl-Dickstein 
http://arxiv.org/abs/1409.5191



- Flatland inspired: exploring the properties of RNNs constrained so that they 
can be embedded in a plane (http://en.wikipedia.org/wiki/Graph_embedding). 
Connections between neurons in a 2D space cannot cross each other. 
What sort of computations are possible when a network is constrained like this?


- Investigate network implementations of chess engine. 
Modern chess engines use game trees and heuristics to play - very different from how humans play the game. 
I'd like to create an engine that works in a more neurobiologically plausible manner. 
This would probably involve unsupervised learning, sampling RNNs, and a feedforward classifying layer or two. 


Let's have a chat if any of these ideas sound interesting to you