Unsupervised Pre-Training of Image Features on Non-Curated Data

Background

See the author’s previous paper which introduced DeepCluster.

Research Goal

Use self-supervised training and clustering to do unsupervised learning

Approach

Deeper Cluster. Rotate images (set of possible rotations: \(Y\)) and optimize K-Means, where \(Z\) are the cluster assignments

\[\arg \min_{\theta, W} \frac{1}{N}\sum_{n=1}^N l(y_n \otimes z_n, W f_{\theta}(x_n))\]

To make the space \(Z \times Y\) smaller, they use 2-level hierarchy of cluster labels. Use 2-level hierarchy. To get superclusters, they split the dataset into \(m\) subsets.

Results

Accuracy of linear classifiers on ImageNet and Places205 using the activations from different layers as fea- tures. We compare a VGG-16 trained with supervision on ImageNet to VGG-16 trained with either RotNet or Deep- erCluster on YFCC100M. Exact numbers are in Appendix