README.md

nnsrm-neurips18

This is the repo for ...

Lu, Q., Chen, P.H., Pillow, J. W., Ramadge, P. J., Norman, K. A., & Hasson, U. (2018). 
Shared Representational Geometry Across Neural Networks. arXiv [cs.LG]. 
Retrieved from http://arxiv.org/abs/1811.11684

What is this project about?

1 sentence summary: different neural networks with the same learning experience acquire representations of the same "shape"

• Here's a short tutorial that describes the minimal analysis pipeline:

Here's a figure showing the activity trajectories from 5 resnets, when they are viewing the same sequence of images.

• Fig left: before alignment; activity trajectories in native spaces
• Fig right: after alignment; it is clear that the geometry of their representations are highly similar
• Here're more animations.

Doc:

Files for the simulation (under simulation/):

• run_sim.ipynb: run the simulation described in the paper
• data_gen.py: make toy data set to train NNs
• models.py: define a simple neural network

Files for the experiments (Files under the root dir):

*The notebooks are not runnable yet, since they depend on some pre-computed data. I'm working on an easy way of hosting the data publicly. Though re-running the whole analysis should be possible.

• show_*.ipynb: load some pre-computed data (e.g. activity from some pre-trained neural networks), apply certain analyses (e.g. SRM), then plot the results
• train_*.py: train some models (e.g. conv nets) on some dataset (e.g. cifar10)
• save_acts_cifar.py: test and save neural network activity matrices
• run_analyses.py: run SRM, RSA, etc.
• models.py: some models (e.g. conv nets)
• resnet.py: resnets from raghakot/keras-resnet [2]
• config.py: define some constants, such as how to re-arrange the ordering of the images in cifar
• data_loader.py: util for loading data

Other files:

• qmvpa: contains some analyses util functions

Dependencies/References:

[1] philipperemy/keract
[2] raghakot/keras-resnet
[3] hypertools
[4] BrainIAK