Quantifying Local Specialization in Deep Neural Networks

Shlomi Hod*, Stephen Casper*, Daniel Filan*, Cody Wild, Andrew Critch, Stuart Russell

A link to the paper will be provided as soon as it is available on arXiv. Check again in a couple of days.

Research Environment Setup

Docker

Useful: Lifecycle of Docker Container

Building the image (done ONCE per machine)

Clone the repository and change to the devops directory.

docker build -t humancompatibleai/nn-clustering .

Creating a container (done ONCE by one user per one machine)

In other words, if you run this commands on, for example, svm - you don't need to do so again. You're supposed then to have a container there already.

First, you need a port number to your Jupyter Netbook - pick up a random number (with your favoriate generator) in the range 8000-8500. We pick up a random number so you don't collide with existing notebooks on that machine.

First run:

• Remove the comments before, and
• Replace <PORT NUMBER> with your random port number (also in the instructions that will come later)

docker run \
-it \
-p <PORT NUMBER>:8888 \
--rm \
--name nn_clustering-$(whoami) \
--runtime=nvidia \  # REMOVE, if you don't have GPU
--mount type=bind,source=/scratch,target=/scratch \  # REMOVE, if not on perceptron or svm machines
humancompatibleai/nn-clustering:latest \
bash

And then type

bash build.sh

NB: to leave the container, use ctrl-P ctrl-Q. Typing exit will destroy the container.

Running the container

docker exec \
-it nn_clustering-$(whoami) \
bash

Coding Environment

Requirements: Python 3.7 (not tested with earlier versions)

The environment is set up in a Python virtual environment. To do so:

1. Clone this repository

2. Install graphviz

   1. Ubuntu/Debian: apt intall graphviz
   2. MacOS: brew install graphviz

3. Install with pipenv install --dev

4. On MacOS only, you will need to install pygraphviz separatly: pipenv run pip install pygraphviz --install-option="--include-path=/usr/local/Cellar/

5. To set up the dependencies and finish, type cd nn_clustering and ./build.sh

6. Install packages with pipenv install --system. Additionally, post-hoc install: pipenv install image-classifiers==1.0.0 tensorflow-datasets

Instructions

This requires the Imagenet2012 validation dataset. You will need to register at http://www.image-net.org/download-images and obtain a link to download the dataset. Then execute the following.

mkdir datasets
mkdir datasets/imagenet2012
cd datasets/imagenet2012
wget [YOUR LINK HERE]

That's all! When running experiments with imagenet models, tfrecords will be created automatically from the .tar file.

See shells/prepare_all.sh for commands to make datasets, train networks, cluster, and perform experiments. We use make with a Makefile to streamline running things. But rather than simply running shells/prepare_all.sh, you may wish to only parts of it. It may take a long time.