Thursday 30 Nov 2017: Dynamics Seminar: An oscillatory neural network model with winner-take-all dynamics explains reaction times in visual search experiments
Roman Borisyuk - Plymouth
LSI Seminar Room A 14:30-15:30
Winner-take-all (WTA) is a computational principle used in cognitive modelling to implement such functions as competitive learning, decision-making, action selection, attention, etc. We suggest a new approach to the WTA modelling that is based on synchronization in an oscillatory network with a central element. The central oscillator (CO) is connected with a set of the so-called peripheral oscillators (POs) by feedforward and feedback connections. We prove that there is a possibility to organize a competition between POs for the synchronization with the CO in such a way that only one PO can win.
We use this WTA oscillatory network to model the results of visual search experiments: POs correspond to different objects in the display and the CO plays the role of the central executive of the attention system. We assume that the strength of the connection from the "target" PO to the CO is higher than the strengths of connections to the CO from other POs, corresponding to distractors. This assumption increases the probability for the "target" PO to win the competition for the synchronization with the CO. The result of model simulations depends on the randomly selected initial values: in some cases these values are in the basin of attraction of WTA dynamics for the target but sometimes they are not. We show that the model correctly reproduces reaction times in visual search tasks of various complexities. Reaction times linearly depend on the number of objects in the display, which is in agreement with experimental evidence. This linearity was not a priori included in the model design; it appeared as a remarkable result of model dynamical properties.