Home

Welcome to the project wiki!

In this project, we are exploring how a scalable higher order objective function for binary networks (such as Hopfield networks) can be used to denoise neural spike data. This essentially becomes a compression algorithm for spike trains, providing researchers with another tool to analyse neural data from the perspective that neurons use precise spike times to encode information.

The abstract for this project can be found here.

In our experiments, we used an in vivo cat dataset which you can find here and an in vitro rat dataset which can be found here. You may need to sign up on those websites before you can get access.

The idea is that given binary patterns, we want to find cluster representations that minimize some distance from the cluster representation (a binary matrix) and the patterns belonging to that cluster. The algorithm we used is the binary Hopfield Network dynamics which can be extended to the higher order form (which we would like to call McCulloch Pitts networks), which we fitted to the data using an objective function called Minimum Energy Flow (MEF) which is detailed here. The library we used that implemented this can be found (here)[https://github.com/TenzinCHW/MEFK.jl].

The experimental steps we used were to first fit the model to the dataset, then perform the dynamics on the same dataset (to convergence) to obtain a distribution over the cluster representations (which is hopefully fewer in number than the original dataset).

We can then ask questions about this distribution or model the sequence of clusters as was done with a Hidden Markov Model (here)[https://redwood.berkeley.edu/wp-content/uploads/2018/01/effenberger2015discovery.pdf].

If you have any feedback or comments, kindly email [email protected], thank you!