Retinal Neural Code Decoding

Unsupervised Learning from Retinal Ganglion Cells

The retina is a complex nervous system which encodes visual stimuli before higher order processing occurs in the visual cortex. Multi Electrode Array Technology is a novel technique that allows to record simoultaneously the activities of thousands of neurons.

The development of methods to make sense of such high dimensional output is an open research area. The goal of our research is to develop computational methods to extract information from Retinal Ganglion Cell neural data. In particular, we are interested in evaluating if we can retrieve the information related to the stimuli proposed to the retina from the RGC firing rate distribution.

To this end, Restricted Boltzmann Machines were applied in order to extract spatio-temporal features from population activity. RBMs are latent variable models capable of modeling a joint distribution and retrieve interesting properties of the input through hidden variables. Specifically, mean-covariance RBMs and conditional RBMs were exploited.

Several dataset were acquired to test the models, with different stimuli shown to the retina: gratings, natural scenes and human actions. Through the grating experiment, we validated the capability of the model to retrieve the regularities associated to different stimuli from the population activity, and the ability to extract also interesting biologically plausible features. The experiment with natural scenese was carried out to test more complex stimuli, with the same modalities. In the human action experiments, we showed the Weizmann dataset to the retina, and demonstrated that we can perform action classification using only neural data.

We also carried out two experiments to evaluate the response of our models to physiological changings in the retina: (i) through different stages of development and (ii) pharmacologically causing impairment in retinal circuitry.

 

 

References:

  • R. Volpi, M. Zanotto, D. Sona, V. Murino
    "Unsupervised Learning of Spatio-Temporal Features from Retinal Ganglion Cells"
    NIPS Workshops - Brains and Bits: Neuroscience Meets Machine Learning, 2016

  • M. Zanotto, S. Di Marco, A. Maccione, V. Murino
    "Modelling Retinal Activity with Restricted Boltzmann Machines: a Study on the Inhibitory Circuitry"
    Computational and Systems Neuroscience (COSYNE), 2015