Representational Dissimilarity Analysis

Representational Dissimilarity Analysis looks at the pairwise correlation of neural activity given different stimuli. How active are particular neurons of a subject if the subject is seeing a house versus if the subject is seeing an apple?

I explored representational dissimilarity of stimuli in medial temporal lobe and deep neural networks as part of my MSNE research project with Prof. Jakob Macke@CNE.

This package resulted from our project and provides automated representational dissimilarity analysis and comparison for preprocessed brain recordings, AlexNet, VGG, and ResNet.

Example usage:

We want to know how 10 images from 10 semantically different classes from Imagenet are represented in a AlexNet.

Here are the images (one column corresponds to one class):

And here's the code:

experiment = ImagenetExperiment()   # init experiment, links to the images
visualization = RDMVis(experiment)  # init plotter

dataset = RDMDataset(experiment)    # this is a pytorch dataset
activations = DNNActivations("alexnet", dataset) # automatically downloads the pretrained alexnet from pytorch

rdms = RDM(activations) 

vis.plot(rdms) 

Giving these results:

Note: Cleaning the code is still work in progress. Example notebooks should be available soon. Feel free to contact me.