L2C: Learning to Cluster

A clustering strategy with deep neural networks. This blog article provides a generic overview.

Introduction

This repository provides the PyTorch implementation of the transfer learning schemes (L2C) and two learning criteria useful for deep clustering:

Meta Classification Likelihood (MCL)* - News: Accepted to ICLR2019 (Title: "Multi-class classification without multi-class labels").
KLD-based Contrastive Loss (KCL)

_{^{*It is renamed from CCL}}

This repository covers following references:

@inproceedings{Hsu19_MCL,
    title =     {Multi-class classification without multi-class labels},
    author =    {Yen-Chang Hsu, Zhaoyang Lv, Joel Schlosser, Phillip Odom, Zsolt Kira},
    booktitle = {International Conference on Learning Representations (ICLR)},
    year =      {2019},
    url =       {https://openreview.net/forum?id=SJzR2iRcK7}
}

@inproceedings{Hsu18_L2C,
    title =     {Learning to cluster in order to transfer across domains and tasks},
    author =    {Yen-Chang Hsu and Zhaoyang Lv and Zsolt Kira},
    booktitle = {International Conference on Learning Representations (ICLR)},
    year =      {2018},
    url =       {https://openreview.net/forum?id=ByRWCqvT-}
}

@inproceedings{Hsu16_KCL,
    title =     {Neural network-based clustering using pairwise constraints},
    author =    {Yen-Chang Hsu and Zsolt Kira},
    booktitle = {ICLR workshop},
    year =      {2016},
    url =       {https://arxiv.org/abs/1511.06321}
}

Preparation

This repository supports PyTorch 1.0, python 2.7, 3.6, and 3.7.

pip install -r requirements.txt

Demo

Supervised Classification/Clustering with only pairwise similarity

# A quick trial:
python demo.py  # Default Dataset:MNIST, Network:LeNet, Loss:MCL
python demo.py --loss KCL

# Lookup available options:
python demo.py -h

# For more examples:
./scripts/exp_supervised_MCL_vs_KCL.sh

Unsupervised Clustering (Cross-task Transfer Learning)

# Learn the Similarity Prediction Network (SPN) with Omniglot_background and then transfer to the 20 alphabets in Omniglot_evaluation.
# Default loss is MCL with an unknown number of clusters (Set a large cluster number, i.e., k=100)
# It takes about half an hour to finish.
python demo_omniglot_transfer.py

# An example of using KCL and set k=gt_#cluster
python demo_omniglot_transfer.py --loss KCL --num_cluster -1

# Lookup available options:
python demo_omniglot_transfer.py -h

# Other examples:
./scripts/exp_unsupervised_transfer_Omniglot.sh

Notes

The clustering results are highly dependent on the performance of the Similarity Prediction Network (SPN). For making a fair comparison, the SPN must be kept the same. Our script trains an SPN with random initialization and random data sampling. Once the SPN model is trained, the script will reuse the saved SPN and avoid training a new one.
The table below presents the clustering performance with the reference SPN [download]. Put the model file into /outputs folder and run demo_omniglot_transfer.py directly to generate the "MCL(k=100)" column.
The performance metric is clustering accuracy (for details, please see L2C paper). Each value in the table is the average of 3 clustering runs. This repository reuses most of the utilities in PyTorch and is different from the Lua-based implementation used in the reference papers. The result (the row with "--Average--") shows the same trend as the papers, but the absolute values have a mild difference. The MCL results here are better than the paper.

Dataset	gt #class	KCL (k=100)	MCL (k=100)	KCL (k=gt)	MCL (k=gt)
Angelic	20	73.2%	82.2%	89.0%	91.7%
Atemayar_Qelisayer	26	73.3%	89.2%	82.5%	86.0%
Atlantean	26	65.5%	83.3%	89.4%	93.5%
Aurek_Besh	26	88.4%	92.8%	91.5%	92.4%
Avesta	26	79.0%	85.8%	85.4%	86.1%
Ge_ez	26	77.1%	84.0%	85.4%	86.6%
Glagolitic	45	83.9%	85.3%	84.9%	87.4%
Gurmukhi	45	78.8%	78.7%	77.0%	78.0%
Kannada	41	64.6%	81.1%	73.3%	81.2%
Keble	26	91.4%	95.1%	94.7%	94.3%
Malayalam	47	73.5%	75.0%	72.7%	73.0%
Manipuri	40	82.8%	81.2%	85.8%	81.5%
Mongolian	30	84.7%	89.0%	88.3%	90.2%
Old_Church_Slavonic_Cyrillic	45	89.9%	90.7%	88.7%	89.8%
Oriya	46	56.5%	73.4%	63.2%	75.3%
Sylheti	28	61.8%	68.2%	69.8%	80.6%
Syriac_Serto	23	72.1%	82.0%	85.8%	89.8%
Tengwar	25	67.7%	76.4%	82.5%	85.5%
Tibetan	42	81.8%	80.2%	84.3%	81.9%
ULOG	26	53.3%	77.1%	73.0%	89.1%
--Average--		75.0%	82.5%	82.4%	85.7%

Compare MCL and KCL

The loss surface of MCL is more similar to the cross-entropy (CE) than KCL. Empirically, MCL converged faster than KCL. For details, please refer to the ICLR paper.

Related Applications

Lane detection for autonomous driving / Instance segmentation

@article{Hsu18_InsSeg,
    title =     {Learning to Cluster for Proposal-Free Instance Segmentation},
    author =    {Yen-Chang Hsu, Zheng Xu, Zsolt Kira, Jiawei Huang},
    booktitle = {accepted to the International Joint Conference on Neural Networks (IJCNN)},
    year =      {2018},
    url =       {https://arxiv.org/abs/1803.06459}
}

Acknowledgments

This work was supported by the National Science Foundation and National Robotics Initiative (grant # IIS-1426998) and DARPA’s Lifelong Learning Machines (L2M) program, under Cooperative Agreement HR0011-18-2-001.