Skip to content

DeepCluster.md

Latest commit

 

History

History
55 lines (49 loc) · 2.92 KB

DeepCluster.md

File metadata and controls

55 lines (49 loc) · 2.92 KB

DeepCluster

  • Title: Deep Clustering for Unsupervised Learning of Visual Features
  • Publication: ECCV, 2018
  • Link: [paper] [code]

Introduce

model

  • When θ is sampled from a Gaussian distribution, without any learning, fθ does not produce good features.
  • However the performance of such random features on standard transfer tasks, is far above the chance level.
  • The last convolutional layer of a random AlexNet achieves 12% in accuracy on ImageNet while the chance is at 0.1%
  • Using this weak signals, the output of the convnet (feature label) is clustered and optimized using them as pseudo-labels.
  • Clustering learns the decision boundary between classes, and if an empty cluster is not considered, assigning all data to one cluster is also regarded as an optimal cluster.
  • Therefore, to prevent this, the empty cluster problem was solved by reallocating the centeriod of an empty cluster to a non-empty cluster.
  • The second problem is trivial parameterization, which is a phenomenon that occurs in the same way with imbalanced data.
  • Avoiding Trivial Solutions.
    • Empty clusters → Reassign cluster
    • Trivial parameterization → Contribute the weight of an input to the loss function by the inverse of the size of its assigned cluster

empty_cluster

Implementation

  • Data Processing
    • Transformation for clustering(resize, crop, horizontal flip, Sobel filter: edge detect)
  • Feature processing
    • PCA(Principal Component Analysis): Dimension reduction, L2-regularization
  • Decide number of clusters(classes): k =10,000
    • Setting up many clusters has a fine-grained grouping effect.

dataprocessing

Experiment Results.

  • Train a linear classifier on top of different frozen convolutional layers.
  • Linear classification on ImageNet and Places using activations from the convolutional layers of an AlexNet as features.
  • On ImageNet, DeepCluster outperforms the state of the art from conv3 to conv5 layers by 3−5%.

experiment

Reference

@article{DBLP:journals/corr/abs-1807-05520,
  author       = {Mathilde Caron and
                  Piotr Bojanowski and
                  Armand Joulin and
                  Matthijs Douze},
  title        = {Deep Clustering for Unsupervised Learning of Visual Features},
  journal      = {CoRR},
  volume       = {abs/1807.05520},
  year         = {2018},
  url          = {http://arxiv.org/abs/1807.05520},
  eprinttype    = {arXiv},
  eprint       = {1807.05520},
  timestamp    = {Mon, 13 Aug 2018 16:46:44 +0200},
  biburl       = {https://dblp.org/rec/journals/corr/abs-1807-05520.bib},
  bibsource    = {dblp computer science bibliography, https://dblp.org}
}