eduzhai > Applied Sciences > Engineering >

End-to-end Learning of Compressible Features

  • king
  • (0) Download
  • 20210506
  • Save

... pages left unread,continue reading

Document pages: 8 pages

Abstract: Pre-trained convolutional neural networks (CNNs) are powerful off-the-shelffeature generators and have been shown to perform very well on a variety oftasks. Unfortunately, the generated features are high dimensional and expensiveto store: potentially hundreds of thousands of floats per example whenprocessing videos. Traditional entropy based lossless compression methods areof little help as they do not yield desired level of compression, while generalpurpose lossy compression methods based on energy compaction (e.g. PCA followedby quantization and entropy coding) are sub-optimal, as they are not tuned totask specific objective. We propose a learned method that jointly optimizes forcompressibility along with the task objective for learning the features. Theplug-in nature of our method makes it straight-forward to integrate with anytarget objective and trade-off against compressibility. We present results onmultiple benchmarks and demonstrate that our method produces features that arean order of magnitude more compressible, while having a regularization effectthat leads to a consistent improvement in accuracy.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×