Geodesics of learned representations

Olivier J. Hénaff, Eero Simoncelli

Research output: Contribution to conferencePaper

Abstract

We develop a new method for visualizing and refining the invariances of learned representations. Specifically, we test for a general form of invariance, linearization, in which the action of a transformation is confined to a low-dimensional subspace. Given two reference images (typically, differing by some transformation), we synthesize a sequence of images lying on a path between them that is of minimal length in the space of the representation (a “representational geodesic”). If the transformation relating the two reference images is linearized by the representation, this sequence should follow the gradual evolution of this transformation. We use this method to assess the invariance properties of a state-of-the-art image classification network and find that geodesics generated for image pairs differing by translation, rotation, and dilation do not evolve according to their associated transformations. Our method also suggests a remedy for these failures, and following this prescription, we show that the modified representation is able to linearize a variety of geometric image transformations.

Original languageEnglish (US)
StatePublished - Jan 1 2016
Event4th International Conference on Learning Representations, ICLR 2016 - San Juan, Puerto Rico
Duration: May 2 2016May 4 2016

Conference

Conference4th International Conference on Learning Representations, ICLR 2016
CountryPuerto Rico
CitySan Juan
Period5/2/165/4/16

Fingerprint

Invariance
Image classification
Linearization
Refining
remedies
medication

ASJC Scopus subject areas

  • Education
  • Computer Science Applications
  • Linguistics and Language
  • Language and Linguistics

Cite this

Hénaff, O. J., & Simoncelli, E. (2016). Geodesics of learned representations. Paper presented at 4th International Conference on Learning Representations, ICLR 2016, San Juan, Puerto Rico.

Geodesics of learned representations. / Hénaff, Olivier J.; Simoncelli, Eero.

2016. Paper presented at 4th International Conference on Learning Representations, ICLR 2016, San Juan, Puerto Rico.

Research output: Contribution to conferencePaper

Hénaff, OJ & Simoncelli, E 2016, 'Geodesics of learned representations', Paper presented at 4th International Conference on Learning Representations, ICLR 2016, San Juan, Puerto Rico, 5/2/16 - 5/4/16.
Hénaff OJ, Simoncelli E. Geodesics of learned representations. 2016. Paper presented at 4th International Conference on Learning Representations, ICLR 2016, San Juan, Puerto Rico.
Hénaff, Olivier J. ; Simoncelli, Eero. / Geodesics of learned representations. Paper presented at 4th International Conference on Learning Representations, ICLR 2016, San Juan, Puerto Rico.
@conference{95e1ea15769d4d6989fdb86e998c7471,
title = "Geodesics of learned representations",
abstract = "We develop a new method for visualizing and refining the invariances of learned representations. Specifically, we test for a general form of invariance, linearization, in which the action of a transformation is confined to a low-dimensional subspace. Given two reference images (typically, differing by some transformation), we synthesize a sequence of images lying on a path between them that is of minimal length in the space of the representation (a “representational geodesic”). If the transformation relating the two reference images is linearized by the representation, this sequence should follow the gradual evolution of this transformation. We use this method to assess the invariance properties of a state-of-the-art image classification network and find that geodesics generated for image pairs differing by translation, rotation, and dilation do not evolve according to their associated transformations. Our method also suggests a remedy for these failures, and following this prescription, we show that the modified representation is able to linearize a variety of geometric image transformations.",
author = "H{\'e}naff, {Olivier J.} and Eero Simoncelli",
year = "2016",
month = "1",
day = "1",
language = "English (US)",
note = "4th International Conference on Learning Representations, ICLR 2016 ; Conference date: 02-05-2016 Through 04-05-2016",

}

TY - CONF

T1 - Geodesics of learned representations

AU - Hénaff, Olivier J.

AU - Simoncelli, Eero

PY - 2016/1/1

Y1 - 2016/1/1

N2 - We develop a new method for visualizing and refining the invariances of learned representations. Specifically, we test for a general form of invariance, linearization, in which the action of a transformation is confined to a low-dimensional subspace. Given two reference images (typically, differing by some transformation), we synthesize a sequence of images lying on a path between them that is of minimal length in the space of the representation (a “representational geodesic”). If the transformation relating the two reference images is linearized by the representation, this sequence should follow the gradual evolution of this transformation. We use this method to assess the invariance properties of a state-of-the-art image classification network and find that geodesics generated for image pairs differing by translation, rotation, and dilation do not evolve according to their associated transformations. Our method also suggests a remedy for these failures, and following this prescription, we show that the modified representation is able to linearize a variety of geometric image transformations.

AB - We develop a new method for visualizing and refining the invariances of learned representations. Specifically, we test for a general form of invariance, linearization, in which the action of a transformation is confined to a low-dimensional subspace. Given two reference images (typically, differing by some transformation), we synthesize a sequence of images lying on a path between them that is of minimal length in the space of the representation (a “representational geodesic”). If the transformation relating the two reference images is linearized by the representation, this sequence should follow the gradual evolution of this transformation. We use this method to assess the invariance properties of a state-of-the-art image classification network and find that geodesics generated for image pairs differing by translation, rotation, and dilation do not evolve according to their associated transformations. Our method also suggests a remedy for these failures, and following this prescription, we show that the modified representation is able to linearize a variety of geometric image transformations.

UR - http://www.scopus.com/inward/record.url?scp=85047011211&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85047011211&partnerID=8YFLogxK

M3 - Paper

AN - SCOPUS:85047011211

ER -