• Top page
  • Timetable
  • Per session
  • Per presentation
  • How to
  • Meeting Planner



視覚 2
Visual System 2

開催日 2014/9/11
時間 17:00 - 18:00
会場 Room G(303)
Chairperson(s) 七五三木 聡 / Satoshi Shimegi (大阪大学大学院医学系研究科 / Graduate School of Medecine, Osaka University, Japan)
小松 英彦 / Hidehiko Komatsu (自然科学研究機構 生理学研究所 / National Institute for Physiological Sciences, Japan)

Four-layer statistical learning model of natural images that explains tuning properties in V2 and V4

  • O1-G-2-3
  • 細谷 晴夫 / Haruo Hosoya:1,2 Hyvarinen Aapo / Aapo Hyvarinen:3 
  • 1:株式会社 国際電気通信基礎技術研究所 (ATR研究所) / ATR Computational Neurosci. Lab., Japan 2:JSTさきがけ / JST Presto, Japan 3:University of Helsinki, Finland / University of Helsinki, Finland 

In light of tremendous high dimensionality of sensory inputs, a sensible hypothesis is that sensory cortex might employ coding strategy that is optimized to the input statistics stemming from the natural environment. Sparse coding and independent component analysis (ICA) are well-known statistical learning models that have successfully explained various properties of V1. However, it is not clear whether this line of modeling can extend to extrastriate areas and, in particular, direct quantitative comparisons to neurophysiological experiments are rare. Here, we extend existing learning models and investigate the connection to experimental findings on V2 and V4. We trained a four-layer model consisting of a stack of ICA modules, with image patches extracted from natural movies of various types including wild and urban life, sport, documentary, film, etc. After learning, the upper two layers in the model exhibited response properties qualitatively and quantitatively compatible with several major neurophysiological results: 1) Third layer represented subfield orientation integration with a distribution biased to smaller orientation differences consistent with a V2 study by Anzai et al. 2) Third layer further exhibited tuning to angles with response specificity to one componential orientation as in a V2 study by Ito and Komatsu. 3) Fourth layer exhibited tuning properties to position-specific curvatures with a distribution biased to acute convexities as in a V4 study by Pasupathy and Connor. However, inspection of the internal representations revealed that units of these two layers did not quite represent angles or curvatures per se. Rather, the basic structures of most units were fairly regular, combining local orientations in co-linear or parallel ways, and thus elicited strong responses to rather straight contours or artificial textures. Many units, however, also had quite complicated structure in the detail, combining slightly or largely different local orientations, which often explained the angle or curvature selectivities.

This work was supported by JST Presto Program.

Copyright © Neuroscience2014. All Right Reserved.