Current browse context:
cs.LG
Change to browse by:
References & Citations
Computer Science > Machine Learning
Title: Lossy Compression for Lossless Prediction
(Submitted on 21 Jun 2021 (v1), last revised 28 Jan 2022 (this version, v5))
Abstract: Most data is automatically collected and only ever "seen" by algorithms. Yet, data compressors preserve perceptual fidelity rather than just the information needed by algorithms performing downstream tasks. In this paper, we characterize the bit-rate required to ensure high performance on all predictive tasks that are invariant under a set of transformations, such as data augmentations. Based on our theory, we design unsupervised objectives for training neural compressors. Using these objectives, we train a generic image compressor that achieves substantial rate savings (more than $1000\times$ on ImageNet) compared to JPEG on 8 datasets, without decreasing downstream classification performance.
Submission history
From: Yann Dubois [view email][v1] Mon, 21 Jun 2021 01:31:44 GMT (11784kb,D)
[v2] Wed, 7 Jul 2021 14:26:46 GMT (11789kb,D)
[v3] Tue, 17 Aug 2021 14:04:36 GMT (11997kb,D)
[v4] Wed, 6 Oct 2021 04:30:55 GMT (12013kb,D)
[v5] Fri, 28 Jan 2022 08:08:19 GMT (12013kb,D)
Link back to: arXiv, form interface, contact.