Maximal Dependence Capturing as a Principle of Sensory Processing

Abstract

Sensory inputs conveying information about the environment are often noisy and incomplete, yet the brain can achieve remarkable consistency in recognizing objects. Presumably, transforming the varying input patterns into invariant object representations is pivotal for this cognitive robustness. In the classic hierarchical representation framework, early stages of sensory processing utilize independent components of environmental stimuli to ensure efficient information transmission. Representations in subsequent stages are based on increasingly complex receptive fields along a hierarchical network. This framework accurately captures the input structures; however, it is challenging to achieve invariance in representing different appearances of objects. Here we assess theoretical and experimental inconsistencies of the current framework. In its place, we propose that individual neurons encode objects by following the principle of maximal dependence capturing (MDC), which compels each neuron to capture the structural components that contain maximal information about specific objects. We implement the proposition in a computational framework incorporating dimension expansion and sparse coding, which achieves consistent representations of object identities under occlusion, corruption, or high noise conditions. The framework neither requires learning the corrupted forms nor comprises deep network layers. Moreover, it explains various receptive field properties of neurons. Thus, MDC provides a unifying principle for sensory processing.

A Local Hebbian Rule Based Neural Network Model of Invariant Object Representation and Classification

“…Occluded forms show lower specificity but are significantly higher than non-specific representations (dotted line). Images in e and f are from ref 49 .…”

“…This result highlights the network's ability to learn from pure experience and generate consistent representations. Importantly, it achieves prospective robustness, defined as consistently representing input patterns it has never experienced 49 .…”

“…We sought a network architecture that allows individual neurons to capture maximum information about distinct objects and can consistently represent an image 49 , meaning that the network responds with the same activity pattern to an image and its various corrupted forms. We constructed a two-layer network.…”

“…Note that the network only experiences <33% of all possible angles. This, the capacity for invariant representation is prospective, meaning the network does not need to encounter all possible variations to represent them consistently 49 .…”

“…Since synaptic changes are oblivious to other parts of the neural network, global changes must arise from how neurons are connected. Here we present two network configurations that capture dependencies among features belonging to the same object and achieve consistent object representation and classification using solely local learning rules 49 . A discriminative network layout produces unique representations of objects robust against signal corruption and occlusion.…”