Low Dimensional Embedding of Neural Responses

NMA-DL (2021) group project: Low Dimensional Embedding of Neural Responses

In this repository, we construct several artifitial neural networks (ANNs) using PyTorch. The goal is to use these seqeunce-to-seqeunce models to extract low-dimensional latent variables from the population Neuropixel data from Steinmetz et al. 2019 Nature.

ANNs:

1. Autoencoder: autoencoder.py
2. Simple GRU: encoder=GRU, decoder=Linear layer; rnn0.py
3. GRU based autoencoder: encoder=GRU, decoder=GRU + Linear layer(final layer); rnnautoencoder.py

Data loading:

1. gen_fake_data.py generates fake Poisson neuronal data, which could be used for initial test of ANNs.
2. get_active_neurons.py could load the raw spiking data and extract active neurons. (We found many silent neurons in the dataset).
3. lfpd.py extracts LFP data.

Results:

In this project, we check the neural data reconstruction errors (MSE) of different state-of-the-art ANNs and traditional PCA (results could be found in the second part of RNN_Notebook.ipynb). We then do logistic regression on extracted latent varibales (independent variables) and behavorial outputs (dependent variable) of each trial and check the classification error of each methods. Classification related code could be found in LOO_CV.ipynb and BehavioralClassifier.ipynb.

We also visualize the latent trajectories and their behavioral representational similarity matrix (RSM), which could be found in the first part of RNN_Notebook.ipynb.

• Note: all the code and results are preliminary and unpublished. If you have any questions, please contact the contributors.