Low-Entropy Latent Variables Hurt Out-of-Distribution Performance

Nandi Schoots; Dylan Cope

Low-Entropy Latent Variables Hurt Out-of-Distribution Performance

Nandi Schoots, Dylan Cope

Informatics

Research output: Chapter in Book/Report/Conference proceeding › Conference paper › peer-review

34 Downloads (Pure)

Abstract

We study the relationship between the entropy of intermediate representations and a model's robustness to distributional shift. We train models consisting of two feed-forward networks end-to-end separated by a discrete n-bit channel on an unsupervised contrastive learning task. Different masking strategies are applied after training that remove a proportion of low-entropy bits, high-entropy bits, or randomly selected bits, and the effects on performance are compared to the baseline accuracy with no mask. We hypothesize that the entropy of a bit serves as a guide to its usefulness out-of-distribution (OOD). Through experiment on three OOD datasets we demonstrate that the removal of low-entropy bits can notably benefit OOD performance. Conversely, we find that top-entropy masking disproportionately harms performance both in-distribution (InD) and OOD.

Original language	English
Title of host publication	International Conference on Learning Representations 2023: Workshop on Domain Generalization
Publication status	Accepted/In press - 2023

Keywords

Out-Of-Distribution Generalisation
Machine Learning
Deep Learning
Entropy
Latent Variables

Access to Document

[Camera Ready] Removing_Low_Entropy_ICLR_WSFinal published version, 1.6 MB

Cite this

@inbook{227eb3e4809e44a793633db2d3ca2719,

title = "Low-Entropy Latent Variables Hurt Out-of-Distribution Performance",

abstract = "We study the relationship between the entropy of intermediate representations and a model's robustness to distributional shift. We train models consisting of two feed-forward networks end-to-end separated by a discrete n-bit channel on an unsupervised contrastive learning task. Different masking strategies are applied after training that remove a proportion of low-entropy bits, high-entropy bits, or randomly selected bits, and the effects on performance are compared to the baseline accuracy with no mask. We hypothesize that the entropy of a bit serves as a guide to its usefulness out-of-distribution (OOD). Through experiment on three OOD datasets we demonstrate that the removal of low-entropy bits can notably benefit OOD performance. Conversely, we find that top-entropy masking disproportionately harms performance both in-distribution (InD) and OOD.",

keywords = "Out-Of-Distribution Generalisation, Machine Learning, Deep Learning, Entropy, Latent Variables",

author = "Nandi Schoots and Dylan Cope",

year = "2023",

language = "English",

booktitle = "International Conference on Learning Representations 2023: Workshop on Domain Generalization",

}

TY - CHAP

T1 - Low-Entropy Latent Variables Hurt Out-of-Distribution Performance

AU - Schoots, Nandi

AU - Cope, Dylan

PY - 2023

Y1 - 2023

N2 - We study the relationship between the entropy of intermediate representations and a model's robustness to distributional shift. We train models consisting of two feed-forward networks end-to-end separated by a discrete n-bit channel on an unsupervised contrastive learning task. Different masking strategies are applied after training that remove a proportion of low-entropy bits, high-entropy bits, or randomly selected bits, and the effects on performance are compared to the baseline accuracy with no mask. We hypothesize that the entropy of a bit serves as a guide to its usefulness out-of-distribution (OOD). Through experiment on three OOD datasets we demonstrate that the removal of low-entropy bits can notably benefit OOD performance. Conversely, we find that top-entropy masking disproportionately harms performance both in-distribution (InD) and OOD.

AB - We study the relationship between the entropy of intermediate representations and a model's robustness to distributional shift. We train models consisting of two feed-forward networks end-to-end separated by a discrete n-bit channel on an unsupervised contrastive learning task. Different masking strategies are applied after training that remove a proportion of low-entropy bits, high-entropy bits, or randomly selected bits, and the effects on performance are compared to the baseline accuracy with no mask. We hypothesize that the entropy of a bit serves as a guide to its usefulness out-of-distribution (OOD). Through experiment on three OOD datasets we demonstrate that the removal of low-entropy bits can notably benefit OOD performance. Conversely, we find that top-entropy masking disproportionately harms performance both in-distribution (InD) and OOD.

KW - Out-Of-Distribution Generalisation

KW - Machine Learning

KW - Deep Learning

KW - Entropy

KW - Latent Variables

M3 - Conference paper

BT - International Conference on Learning Representations 2023: Workshop on Domain Generalization

ER -

Low-Entropy Latent Variables Hurt Out-of-Distribution Performance

Abstract

Keywords

Access to Document

Fingerprint

Cite this