Laboratory of Physiology of Cognitive Processes
2008
  • Title:Decoding kinetic depth using only the temporal structure of spike trains from area MT
  • Authors:N. Tsuchiya; A. Maier; N. K. Logothetis; D. A. Leopold
  • Title of Journal:38th Annual Meeting of the Society for Neuroscience (Neuroscience 2008)
  • Year:2008
  • DOI:
Abstract
What kind of neuronal code is correlated with conscious perception? This question has been frequently asked by recording neuronal activity from various areas of the visual cortex of monkeys who reported their percept upon seeing ambiguous stimuli, such as binocular rivalry, random dot motion and structure-from-motion. The firing rate code is likely to be utilized in the central nervous system, and, critically, it is known to be correlated with perception; the number of the spikes emitted from a single neuron over 1 second has been shown to be able to predict monkeys perceptual decision above chance (Leopold & Logothetis 1996, Britten et al 1996, Bradley et al. 1998). Another candidate is the temporal code, which has not been systematically studied with ambiguous stimulation. Here we examined the extent to which the temporal structure in spikes is correlated with a reported percept. We trained two monkeys to report perceived direction of rotation while they viewed a structure-from-motion stimulus. We manipulated the ambiguity of the stimuli via interocular disparity. We recorded neuronal activity from the motion sensitive area MT with 8-10 microelectrodes at a time. We estimated the spike spectrum using a 500 msec window. For each 500 msec window, we divided the spectrum by the mean firing rate to obtain a rate-normalized spectrum, which emphasizes changes in temporal structure that are independent of changes in the firing rate (Pesaran et al 2002). We found that a strong temporal signature of perceived direction in the unambiguous trials. In ambiguous trials, the effect was weaker, but nonetheless present for several neurons. To quantify the reliability of the observed modulation, we used a decoding approach. 70% of the trials were used as a training set to train a regularized least square classifier, and the rest of