NNclustering

The demo code for "Neural network-based clustering using pairwise constraints" (http://arxiv.org/abs/1511.06321)

Prerequest

Install the hungarian algorithm for optimal cluster assignment.

MNIST

Demo

The demo code is modified from szagoruyko's work.

Do classification on MNIST:

th demo.lua

Do clustering on MNIST:

th demo.lua --clustering 1 -b 32 -r 0.01

To lookup the available arguments:

th demo.lua -h

Usage of BatchKLDivCriterion layer

It can handle 3 types of 'target'. The notation of n is the size of mini-batch.

1. nx1 class label: It uses the same label as classification task. The pairwise relationship will be enumerated in BatchKLDivCriterion automatically.
2. nxn relationship matrix: target[i][j]= 1:similar pair; -1:dissimilar pair; 0:no relationship
3. nx3 tuple: (i, j, relationship) where i and j are the index of sample inside the mini-batch and relationship indicates similar/dissimilar pair. relationship= 1:similar pair; 0: dissimilar pair

Acknowledgments

This work was supported by the National Science Foundation and National Robotics Initiative (grant # IIS-1426998).