Anticipative Dynamics of a Retina studied by Machine Learning

YuTing Huang^1*, Kuan-Han Chen¹, Qi-Rong Lin¹, C.K. Chan¹

¹Institute of Physics, Academia Sinica, Taipei, Taiwan

* Presenter:YuTing Huang, email:bloodsuckant@gmail.com

Anticipative dynamics (AD) of a retina is its ability to predict future events of incoming signals.The mechanism of AD is still unknown. Retinal responses of a time dependent input I(x, y, t) in the form of spike trains are studied by three methods to understand the encoding and anticipative mechanism of a retina. The first is the use of bull frog’s retina to produce spike trains recorded by a multi-electrode array (MEA). The second is to generate spike trains from a physiological realistic simulation known as COREM by using the same input as I(x, y, t) in the real retina. The third one is to use machine learning to train a CNN to mimic the spike trains from the MEA and from COREM. Our goal is to investigate what kind of structures in the retinal network is needed to produce anticipative dynamics. In this year’s project, we have successfully implemented the needed code for COREM and the CNN. Preliminary results show that the CNN can reveal some characteristic of the simulated retina such as the receptor field. However, more works are needed to use these new tools for the understanding of our experimental data.

Keywords: Anticipative Dynamics, Retina, Machine Learning