Abstract
Biological motion recognition can show perceptual multi-stability with spontaneous transitions between the perceived walking directions (Vanrie, 2004). It is also subject to adaptation, as demonstrated by high-level after-effects and fMRI repetition suppression in relevant areas. Existing neural models for action recognition do not account for these phenomena. We present a neural mass model that accounts for these effects, and which permits to study the dynamic interplay between multi-stability and adaptation. METHODS: The model is based on a 2D dynamic neural network (field) that is composed from ‘snapshot neurons’ that encoding different views of body postures, which arise during the recognized action. These neurons are laterally coupled, resulting in competition between different views, and in sequence selectivity with respect to the temporal order of the presented stimulus frames. The model neurons are noisy, and include two different adaptation processes: a) a firing rate (FR) fatigue process, which increases their thresholds after firing, and b) an input fatigue (IF) process, which reduces the efficiency of input synapses after frequent stimulation. Parameters of the model were fitted to data from action- and shape-selective neurons in the STS and area IT of macaques. RESULTS /DISCUSSION: The neural field model reproduces the multi-stability of biological motion perception and spontaneous perceptual switching. The model also reproduces in detail neurophysiological data on adaptation effects in area IT, and specifically confirms that input fatigue is critical to reproduce the relevant data (deBaene, 2011). Using the same adaptation mechanism in a model for the recognition of body motion, results in adaptation effects that depend on the relative dominance of IF vs. FR fatigue. For dominance of FR fatigue the model reproduces electrophysiological observations showing very weak adaptation in action-selective neurons (Caggiano, 2012, Kilner, 2013).
Meeting abstract presented at VSS 2015