SENSORIUM+ 2022 Competition Submission

Contents

1. Overview
2. Setup Instructions
3. Changes Implemented
4. Evaluation

1. Overview

The NeurIPS 2022 The SENSORIUM competition aimed to find the best neural predictive model that can predict the activity of thousands of neurons in the primary visual cortex of mice in response to natural images.

Tracks:

SENSORIUM - Stimulus-only - Assessed on how well they predict neural activity solely in response to the visual stimulus averaged over all trials.
SENSORIUM+ - Stimulus-and-Behavior - Assessed based on how well they can predict neural activity given additional behavioral variables.

This repository contains our submission for this competition, where we attempted to improve the baseline model for the competition track- Sensorium+.

2. Setup Instructions

Below is a step-by-step guide for getting started.

1. Pre-requisites

• install docker and docker-compose
• install git
• clone the repo via git clone https://github.com/sinzlab/sensorium.git

2. Download neural datasets

You can download the data from https://gin.g-node.org/cajal/Sensorium2022 and place it in sensorium/notebooks/data.
Note: Downloading all the files at once as a directory leads to subsequent errors. Hence, download all datasets individually.

3. Run the example notebooks

Start Jupyter environment

cd sensorium/
docker-compose run -d -p 10101:8888 jupyterlab

Now, type localhost:10101 in your browser address bar, and you are good to go!

3. Changes implemented:

• While finding the mean and variance of the neuron specific receptive field, we experimented with introducing multiple samplings of the same to increase number of parameters.
  
• Also experimented with tuning the hyper-parameters such as making changes in the batch sizes of the datasets during training of the model, as well as layers, hidden channels, learning rate.

4. Evaluation

• We use the predictions stored in the .csv submission file, generated using the API provided by Sensorium to evaluate the performance of our method.
• We were able to replicate the Single Trial Correlation and The Correlation Average score of the State-Of-The-Art Model provided.

Tutorial Notebooks

Dataset tutorial: Visual rendering and analysis of the structure of datasets through graphs and tables, and the code to turn it into a PyTorch DataLoader.
Model tutorial: How to train and evaluate baseline models.
Submission tutorial: Use the Sensorium API to generate a .csv file and make a submission to the competition.

Team members

Manav Goyal*, Manyana Tiwari*, Sarthak Gupta*