Home

Awesome

Information Maximizing Self Augmented Training (IMSAT)

This is a reproducing code for IMSAT [1]. IMSAT is a method for discrete representation learning using deep neural networks. It can be applied to clustering and hash learning to achieve the state-of-the-art results. This is the work performed while Weihua Hu was interning at Preferred Networks.

Requirements

You must have the following already installed on your system.

Quick start

For reproducing the experiments on MNIST datasets in [1], run the following codes.

calculate_distance.py can be used to calculate the perturbation range for Virtual Adversarial Training [2]. For MNIST dataset, we have already calculated the range.

Datasets

All the datasets used in the paper can be downloaded here.

Reference

[1] Weihua Hu, Takeru Miyato, Seiya Tokui, Eiichi Matsumoto and Masashi Sugiyama. Learning Discrete Representations via Information Maximizing Self-Augmented Training. In ICML, 2017. Available at http://arxiv.org/abs/1702.08720

[2] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, Ken Nakae, and Shin Ishii. Distributional smoothing with virtual adversarial training. In ICLR, 2016.