NatHacks22NeuralSignalDecoding

The goal of this project is using PyTorch to process mouse brain neuron signal data in order to train a model that automatically infers which type of sound is heard by the subject when an external sound is being played to the subject (where the sound has to be one of the 10 sounds in the training dataset).

Model Structure

The model used to learn the classification of sound being played is simple, consisting of a pooling layer to reduce the input size (as in the provided dataset each trial has 85 neurons activities recorded) and a linear regression classifier. The data for training and testing will be loaded to GPU memory and processed there if you have a NVIDIA GPU available. The model obtains 94% accuracy on reserved test set of 200 trials.

What I learned

In this project I learned about how to use machine learning to process actual experimental data, and especially handling the problem of overfitting when there are not many data samples available.

Acknowledgements

I would acknowledge that the PyTorch webpage tutorial has helped me a lot during writing this program, as I've only picked up this as my first ML library and learned most of the implementation from the tutorial.

I would also like to thank NatHacks for hosting workshops on basics of ML, including introduction to techniques used for supervised/unsupervised learning, which has inspired me to try out a few other approaches to this problem besides linear regression.

Built With

• matlab
• numpy-and-scipy
• python
• pytorch