Parvathy Neelakandan: The formation of positive prediction-error neurons in heterogeneous interneuron circuits

BCCN Berlin / Technische Universität Berlin

Abstract

Self-generated and externally generated actions stimulate our sensory systems. However, perceptual systems effectively differentiate external and self-generated actions and perceive them accurately. Predictive processing, a popular theory that takes into account this distinction states that perceptual systems are constantly involved in making predictions of sensory signals and comparing them with actual sensory signals. Deviations of predicted signals from sensory signals, called prediction errors then indicate the changes in the external world that are not recognized by internal predictions. This error may be adjusted to eliminate future discrepancies. Theoretically, the deviation of predictions from sensory input can be negative or positive, depending on how strong or weak the predictions are, and can be referred to as negative prediction errors and positive prediction errors. Neurons encoding these prediction errors are called negative and positive prediction error neurons depending on the error they encode. There are only few computational models that examine the circuit-level formation of neurons encoding prediction errors based on predictive processing and sensorimotor integration in V1. A recent work has addressed the study of negative prediction error neurons in a circuit model comprising different interneuron types. Here, we expand on this work by studying the formation of positive prediction error neurons in a simple heterogeneous interneuron circuit with PV, SST, and VIP neurons. Different input configurations for SST and VIP neurons, as well as simulated optogenetic manipulations, demonstrate the possibilities for the formation of positive prediction error neurons in the circuit. Positive prediction error neurons are formed by the balance of distinct pathways of excitation and inhibition to the excitatory neurons. An inhibitory plasticity rule and a backpropagation of error rule on four different plastic weights in the model generates a circuit of positive prediction error neurons that responds only when the predictions are weaker than the sensory input.

 

Additional Information

Master Thesis Defense

 

Organized by

Prof. Henning Sprekeler   & Prof. Susanne Schreiber   / Lisa Velenosi