Behavioral benefits of spatial attention explained by multiplicative gain, not receptive field shifts, in a neural network model.

This repository contains the implementation of attention models in: Kai Fox, Dan Birman, and Justin Gardner. “Behavioral benefits of spatial attention explained by multiplicative gain, not receptive field shifts, in a neural network model.” bioRxiv. 2022. [preprint].

The main contents are

• code/script/figs/ – Scripts to quickly generate each figure of the manuscript using pre-generated data. The order of the figures as presented in the manuscript is fig-task, fig-cornet, fig-gauss, fig-shift, fig-sensitivity, fig-shrink*, fig-flat, fig-reconstruct (w/ supplements fig-avgsupp* and fig-discrimsupp*). Those listed with * appear only in the revised manuscript.
• code/script/ – Scripts to generate data for figures, intended to be run in a high-performance-computing (HPC) environment.
• code/lib – Classes and methods to support the mechanistic models of attention studied in the figures.
• notebooks/ – Markdown files for each figure describing the use of the HPC scripts to generate data referenced in the corresponding code/script/figs file.

Setup

This repo is built mainly on torch, numpy, matplotlib, and h5py, but a full list of libraries and versions for setup using your virtual environment manager of choice may be found in requirements.txt

Much of the code is designed to be run across a combination of local and high-performance-computing environments, so all paths are specified relative to CODE and DATA environment variables (accessed in python via lib.paths) that should be defined in a bash_rc or other environment startup script.