A convolutional subunit model for neuronal responses in macaque V1

Research output: Contribution to journalArticle

Abstract

The response properties of neurons in the early stages of the visual system can be described using the rectified responses of a set of self-similar, spatially shifted linear filters. In macaque primary visual cortex (V1), simple cell responses can be captured with a single filter, whereas complex cells combine a set of filters, creating position invariance. These filters cannot be estimated using standard methods, such as spike-triggered averaging. Subspace methods like spike-triggered covariance can recover multiple filters but require substantial amounts of data, and recover an orthogonal basis for the subspace in which the filters reside, rather than the filters themselves. Here, we assume a linear-nonlinear-linear-nonlinear (LN-LN) cascade model in which the first LN stage consists of shifted (“convolutional”) copies of a single filter, followed by a common instantaneous nonlinearity. We refer to these initial LN elements as the “subunits” of the receptive field, and we allow two independent sets of subunits, each with its own filter and nonlinearity. The second linear stage computes a weighted sum of the subunit responses and passes the result through a final instantaneous nonlinearity. We develop a procedure to directly fit this model to electrophysiological data. When fit to data from macaque V1, the subunit model significantly outperforms three alternatives in terms of cross-validated accuracy and efficiency, and provides a robust, biologically plausible account of receptive field structure for all cell types encountered in V1.

Original languageEnglish (US)
Pages (from-to)14829-14841
Number of pages13
JournalJournal of Neuroscience
Volume35
Issue number44
DOIs
StatePublished - Nov 4 2015

Fingerprint

Macaca
Nonlinear Dynamics
Visual Cortex
Neurons

Keywords

  • Model
  • Receptive field
  • Subunits
  • V1

ASJC Scopus subject areas

  • Neuroscience(all)

Cite this

A convolutional subunit model for neuronal responses in macaque V1. / Vintch, Brett; Movshon, J. Anthony; Simoncelli, Eero.

In: Journal of Neuroscience, Vol. 35, No. 44, 04.11.2015, p. 14829-14841.

Research output: Contribution to journalArticle

@article{fbefd0e777a14f81b3c9f0ef93090667,
title = "A convolutional subunit model for neuronal responses in macaque V1",
abstract = "The response properties of neurons in the early stages of the visual system can be described using the rectified responses of a set of self-similar, spatially shifted linear filters. In macaque primary visual cortex (V1), simple cell responses can be captured with a single filter, whereas complex cells combine a set of filters, creating position invariance. These filters cannot be estimated using standard methods, such as spike-triggered averaging. Subspace methods like spike-triggered covariance can recover multiple filters but require substantial amounts of data, and recover an orthogonal basis for the subspace in which the filters reside, rather than the filters themselves. Here, we assume a linear-nonlinear-linear-nonlinear (LN-LN) cascade model in which the first LN stage consists of shifted (“convolutional”) copies of a single filter, followed by a common instantaneous nonlinearity. We refer to these initial LN elements as the “subunits” of the receptive field, and we allow two independent sets of subunits, each with its own filter and nonlinearity. The second linear stage computes a weighted sum of the subunit responses and passes the result through a final instantaneous nonlinearity. We develop a procedure to directly fit this model to electrophysiological data. When fit to data from macaque V1, the subunit model significantly outperforms three alternatives in terms of cross-validated accuracy and efficiency, and provides a robust, biologically plausible account of receptive field structure for all cell types encountered in V1.",
keywords = "Model, Receptive field, Subunits, V1",
author = "Brett Vintch and Movshon, {J. Anthony} and Eero Simoncelli",
year = "2015",
month = "11",
day = "4",
doi = "10.1523/JNEUROSCI.2815-13.2015",
language = "English (US)",
volume = "35",
pages = "14829--14841",
journal = "Journal of Neuroscience",
issn = "0270-6474",
publisher = "Society for Neuroscience",
number = "44",

}

TY - JOUR

T1 - A convolutional subunit model for neuronal responses in macaque V1

AU - Vintch, Brett

AU - Movshon, J. Anthony

AU - Simoncelli, Eero

PY - 2015/11/4

Y1 - 2015/11/4

N2 - The response properties of neurons in the early stages of the visual system can be described using the rectified responses of a set of self-similar, spatially shifted linear filters. In macaque primary visual cortex (V1), simple cell responses can be captured with a single filter, whereas complex cells combine a set of filters, creating position invariance. These filters cannot be estimated using standard methods, such as spike-triggered averaging. Subspace methods like spike-triggered covariance can recover multiple filters but require substantial amounts of data, and recover an orthogonal basis for the subspace in which the filters reside, rather than the filters themselves. Here, we assume a linear-nonlinear-linear-nonlinear (LN-LN) cascade model in which the first LN stage consists of shifted (“convolutional”) copies of a single filter, followed by a common instantaneous nonlinearity. We refer to these initial LN elements as the “subunits” of the receptive field, and we allow two independent sets of subunits, each with its own filter and nonlinearity. The second linear stage computes a weighted sum of the subunit responses and passes the result through a final instantaneous nonlinearity. We develop a procedure to directly fit this model to electrophysiological data. When fit to data from macaque V1, the subunit model significantly outperforms three alternatives in terms of cross-validated accuracy and efficiency, and provides a robust, biologically plausible account of receptive field structure for all cell types encountered in V1.

AB - The response properties of neurons in the early stages of the visual system can be described using the rectified responses of a set of self-similar, spatially shifted linear filters. In macaque primary visual cortex (V1), simple cell responses can be captured with a single filter, whereas complex cells combine a set of filters, creating position invariance. These filters cannot be estimated using standard methods, such as spike-triggered averaging. Subspace methods like spike-triggered covariance can recover multiple filters but require substantial amounts of data, and recover an orthogonal basis for the subspace in which the filters reside, rather than the filters themselves. Here, we assume a linear-nonlinear-linear-nonlinear (LN-LN) cascade model in which the first LN stage consists of shifted (“convolutional”) copies of a single filter, followed by a common instantaneous nonlinearity. We refer to these initial LN elements as the “subunits” of the receptive field, and we allow two independent sets of subunits, each with its own filter and nonlinearity. The second linear stage computes a weighted sum of the subunit responses and passes the result through a final instantaneous nonlinearity. We develop a procedure to directly fit this model to electrophysiological data. When fit to data from macaque V1, the subunit model significantly outperforms three alternatives in terms of cross-validated accuracy and efficiency, and provides a robust, biologically plausible account of receptive field structure for all cell types encountered in V1.

KW - Model

KW - Receptive field

KW - Subunits

KW - V1

UR - http://www.scopus.com/inward/record.url?scp=84946429380&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=84946429380&partnerID=8YFLogxK

U2 - 10.1523/JNEUROSCI.2815-13.2015

DO - 10.1523/JNEUROSCI.2815-13.2015

M3 - Article

C2 - 26538653

AN - SCOPUS:84946429380

VL - 35

SP - 14829

EP - 14841

JO - Journal of Neuroscience

JF - Journal of Neuroscience

SN - 0270-6474

IS - 44

ER -