Purchase this article with an account.
Johannes Burge, Wilson Geisler; Optimal retinal speed estimation in natural image movies. Journal of Vision 2013;13(9):453. doi: 10.1167/13.9.453.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
The neural computations underlying selective perceptual invariance are enormously complex. Many studies of neural encoding-decoding assume neurons with invariant tuning functions. Here, we show how construct neurons that are largely invariant to irrelevant natural image variation. We do so by applying a task-specific encoding-decoding framework that simultaneously specifies how to encode and decode task-relevant information from the retinal image. We use the framework to estimate retinal image speed from photoreceptor responses to natural image movies. The movies were consistent with local translation relative to fronto-parallel and slanted surfaces. The distribution of slants in natural scenes was well-approximated by the distribution of slants in the analysis. The space-time receptive fields (RFs) that optimally encode information relevant for estimating speed are direction selective but, interestingly, they are not speed-tuned. Appropriate non-linear combination of the RF responses yields a new population of neurons that are speed tuned and are (largely) invariant to irrelevant stimulus dimensions. These neurons represent the log-likelihood (LL) of speed and have tuning curves that are approximately log-Gaussian in shape. MAP decoding yields unbiased speed estimates over a wide range (-8 to 8 deg/sec). The optimal space-time RFs and speed-tuned LL neurons share many properties with neurons in cortex. Most motion sensitive neurons in V1 and MT are direction but not speed selective whereas ~25% of V1 and MT neurons are speed tuned (Priebe, Lisberger, Movshon, 2006). Cortical speed-tuned neurons have tuning curves that are log-Gaussian in shape (Nover, Anderson, DeAngelis, 2005). Critically, the optimal space-time RFs and speed-tuned neurons from our analysis were not arbitrarily chosen to match the properties of neurophysiological RFs. Rather, they emerge from a task-specific analysis of natural signals. We find it remarkable that an ideal-observer analysis, with appropriate biological constraints and zero free parameters, predict many of the dominant neurophysiological features of speed processing.
Meeting abstract presented at VSS 2013
This PDF is available to Subscribers Only