This repository contains the code for the paper "Neural networks behave as hash encoders: An empirical study" by Fengxiang He, Shiye Lei, Jianmin Ji, and Dacheng Tao.
-
Model capacity
-
Training time
-
Sample size
- Python3.6
- Tensorflow1.9
- Keras2.2
- MNIST dataset
- CIFAR-10 dataset
- For MNIST:
python run_layer_width_train_mlp.py --dataset mnist --depth 1 --begin_repeat 1 --repeat 10 - For CIFAR-10:
python run_layer_width_train_mlp.py --dataset cifar10 --depth 5 --begin_repeat 1 --repeat 10
- For MNIST:
python run_sample_size_train_mlp.py --dataset mnist --depth 1 --begin_repeat 1 --repeat 10 - For CIFAR-10:
python run_sample_size_train_mlp.py --dataset cifar10 --depth 5 --begin_repeat 1 --repeat 10
- For MNIST:
python run_training_time_train_mlp.py --dataset mnist --depth 1 --begin_repeat 1 --repeat 10 - For CIFAR-10:
python run_training_time_train_mlp.py --dataset cifar10 --depth 5 --begin_repeat 1 --repeat 10
- For MNIST:
python run_layer_width_encoding_properties.py --dataset mnist --depth 1 --begin_repeat 1 --repeat 10 - For CIFAR-10:
python run_layer_width_encoding_properties.py --dataset cifar10 --depth 5 --begin_repeat 1 --repeat 10
- For MNIST:
python run_sample_size_encoding_properties.py --dataset mnist --depth 1 --begin_repeat 1 --repeat 10 - For CIFAR-10:
python run_sample_size_encoding_properties.py --dataset cifar10 --depth 5 --begin_repeat 1 --repeat 10
- For MNIST:
python run_training_time_encoding_properties.py --dataset mnist --depth 1 --begin_repeat 1 --repeat 10 - For CIFAR-10:
python run_training_time_encoding_properties.py --dataset cifar10 --depth 5 --begin_repeat 1 --repeat 10
For the well-trained models in our paper, please kindly contact Shiye Lei at leishiye@gmail.com.
@article{he2021neural,
title={Neural networks behave as hash encoders: An empirical study},
author={He, Fengxiang and Lei, Shiye and Ji, Jianmin and Tao, Dacheng},
journal={arXiv preprint arXiv:2101.05490},
year={2020}
}
For any issue, please kindly contact
Fengxiang He: fengxiang.f.he@gmail.com
Shiye Lei: leishiye@gmail.com
Jianmin Ji: jianmin@ustc.edu.cn
Dacheng Tao: dacheng.tao@sydney.edu.au
--
Last update: Fri 15 Jan 2021