This code generates the analyses in the paper "Structured random receptive fields enable informative sensory encodings". https://www.biorxiv.org/content/10.1101/2021.09.09.459651
Summary: In this work, we describe an approach to model receptive fields of biological neurons as samples from Gaussian Processes (GP). We successfully model receptive fields datasets in two disparate sensory regions, insect mechanosensors and mammalian primary visual cortex (V1). Using our models, we also show a novel initialization scheme that improves the learning efficiency of artificial neural networks (ANNs).
- Do a clean download of the repository. Go inside the repo.
cd \path\to\structured_random_features\ - Make a virtual Conda environment. We will be using python version 3.8.
conda create --name random_features python=3.8 - Activate the new conda environment
conda activate random_features - Run the make file to download all the datasets and preprocess.
make data - Run the notebooks in notebooks/reports folder.
Remark: Always work in the newly created conda environment.
-
The source code is in the /src folder. The jupyter notebooks in /notebooks/reports folder show the analyses used in the paper.
-
Notebooks 1--5 show how we fit our covariance models to biological data. We plot these results in Notebooks 16.0--16.4.
-
Notebooks 6 and 7 show the results on the time-series classification tasks. We plot these results in Notebook 16.5.
-
Notebooks 8--11 show the results on the digit classification tasks. We plot these results in Notebook 16.6.
-
Notebooks 12--15 show our classification performance when initializing networks with structured random features. We plot these results in Notebook 16.7 and 16.8
-
Notebooks 16.xx are used to plot the figures shown in the paper. They assume that you have run the relevant notebooks and saved the results.
Remark: This project structure is based on the cookiecutter data science project template.