NeuCEPT

Source code for the experiments in the paper NeuCEPT: Locally Discover Neural Networks' Mechanism via Critical Neurons Identification with Precision Guarantee. Our method is also named RATIONAL in some experiments.

Directory structure

• mnist: source code for the LeNet model on MNIST
• cifar10: source code for the VGG11 model on CIFAR-10

Dependencies

This codebase has been developed and tested only with python 3.8.10 and on a linux 64-bit operation system.

We recommend using conda to create a development environment with necessary packages. Refer to https://www.anaconda.com/products/individual for instructions on how to install conda.

Set up a conda environment

We have prepared a file containing the same environment specifications that we use. To reproduce this environment (only on a linux 64-bit OS), execute the following command:

$ conda create --name rational --file spec-list.txt

Usage

First, activate the created environment with the following command:

$ conda activate rational

Refer to the README.md files in the mnist and cifar10 directories for more information on reproducing the experiments on the MNIST and CIFAR-10 datasets, respectively.