Increase in Mutual Information During Interaction with the Environment Contributes to Perception

Abstract

Perception and motor interaction with physical surroundings can be analyzed by the changes in probability laws governing two possible outcomes of neuronal activity, namely the presence or absence of spikes (binary states). Perception and motor interaction with the physical environment are partly accounted for by a reduction in entropy within the probability distributions of binary states of neurons in distributed neural circuits, given the knowledge about the characteristics of stimuli in physical surroundings. This reduction in the total entropy of multiple pairs of circuits in networks, by an amount equal to the increase of mutual information, occurs as sensory information is processed successively from lower to higher cortical areas or between different areas at the same hierarchical level, but belonging to different networks. The increase in mutual information is partly accounted for by temporal coupling as well as synaptic connections as proposed by Bahmer and Gupta (Front. Neurosci. 2018). We propose that robust increases in mutual information, measuring the association between the characteristics of sensory inputs' and neural circuits' connectivity patterns, are partly responsible for perception and successful motor interactions with physical surroundings. The increase in mutual information, given the knowledge about environmental sensory stimuli and the type of motor response produced, is responsible for the coupling between action and perception. In addition, the processing of sensory inputs within neural circuits, with no prior knowledge of the occurrence of a sensory stimulus, increases Shannon information. Consequently, the increase in surprise serves to increase the evidence of the sensory model of physical surroundings.

Keywords: Shannon information; affordance; embedded cognitive theory; sparse coding; surprisal; temporal coupling; temporal processing.

Figure 1

This schematic depicts a typical configuration of processing hubs formed by cortical areas in the brain. A large cortical area, A, is shown to interact with several other circuits (B, C, and D) by sending and/or receiving inputs when processing from inputs related to a common stimulus or input source. In a task with a greater complexity, another circuit, E (broken line), is shown to be involved. The addition of another circuit, to the network processing a given stimulus, will decrease the conditional entropy (area of circle A, excluding overlapping regions) and increase the mutual information (area of all overlapping regions within circle A), which will improve the signal to noise ratio (Equations (11)–(13)).

Figure 2

This schematic (adapted from Bahmer and Gupta [21]) depicts the coincidence detection (D) leading to the activation of a set of neurons. This sparse set of neurons are activated (D) when the excitatory phase of low frequency oscillation (B) coincides with gamma synchronized synaptic discharge [27] from ramping neurons (A). Sensory stimuli reset the phase of low-frequency oscillations, leading to a random shift in the excitatory phase (orange shaded area) of the oscillations in relation to the ramping activities, which represent internal states of the brain. Based on the fraction of the excitatory phase of the total cycle length, the excitatory phase of the low-frequency oscillation will coincide with a high-level firing state of a ramping neuron according to a probability value of 1/k. The surprisal or Shannon information from the coincidence detection, which results in the firing of a set of neurons (sparse coding), encodes the sensory and motor interaction with external environment.

Figure 3

This schematic depicts web-like connectivity resulting from an increased joint probability of activity $\left(p\left(x_i,y_j\right)\right)$, (shown as double-headed arrows) of pairs of neurons in the circuits, X and Y. Notice that the web-like configuration in A and B are different, which depicts different consequences due to differences in stimuli, affecting the probability laws that govern neuronal activities. Moreover, these differences in web-like patterns in two scenarios can result in differences in the outputs, which would be responsible for differences in motor or behavioral outcomes.

Figure 4

This schematic depicts interactions between different regions that construct dorsal and ventral streams, which serve as large populations of neurons with binary states, providing a large entropy. A large entropy can allow robust increases in mutual information via interactions with many circuits, helping to improve the signal to noise ratio (Equations (11)–(13)). Also notice that many regions of the dorsal and ventral streams have extensive connections with one another and with the cerebellum.

Figure 5

This schematic illustrates the relative roles of mutual information and Shannon information in action and perception. Mutual information is the increase in certainty in neural circuits given the knowledge about environmental sensory stimuli, and therefore it represents the environmental object properties in perception and action, also called affordance (shown as an empty frame). Since for Shannon information, one assumes that there is no previous knowledge about the sensory stimuli, its increase serves as the evidence of the source of the sensory stimuli (shown as several symbols). Both mutual information (affordance) and evidence of the sensory stimuli are responsible for action and perception. The perception results from the guidance of action by perception (embedded cognition).