ArXiv • SVRHM • NeurIPS 2019 • Notebooks

How Convolutional Neural Network Architecture Biases Learned Opponency and Spectral Tuning

Ethan Harris (@ethanwharris), Daniela Mihai (@Ddaniela13) and Jonathon Hare (@jonhare)

About

This repository contains the code for our paper 'How Convolutional Neural Network Architecture Biases Learned Opponency and Spectral Tuning' and a previous version entitled 'Spatial and Colour Opponency in Anatomically Constrained Deep Networks', accepted to the NeurIPS 2019 workshop on Shared Visual Representations in Humans and Machines (SVRHM).

Notebooks

	Spectral Opponency: Generating the spectral opponency figures from the paper.
	Spatial Opponency: Generating the spatial opponency figures from the paper.
	Double Opponency: Generating the double opponency figures from the paper.
	CIELAB: Experiments in CIELAB colour space.
	Channel Shuffled: Experiments in with random channel shuffling.
	Classification Performance: Accuracy plots from trained models.
	Gratings: Generate example grating images.
	Ventral Depth - Spectral: Plot spectral opponency as a function of ventral depth.
	Ventral Depth - Spatial: Plot spatial opponency as a function of ventral depth.
	Colour Distribution: Plot of most excitatory and inhibitory colours.
	Characterising a Single Cell: Experiments showing the characterisation of a single cell.
	Colour Perception: Experiments on colour sensitivity in Humans and Machines.
	Mouse Spatial Tuning: Spatial tuning curves for cells in the Mouse LGN from Zhao et al.