Heeger presents a simple and lucid framework for a unified theory of cortical function that he suggests should be useful for guiding both neuroscience and artificial intelligence work. I'm passing on the summary, abstract and the first part of the introduction (the article, unfortunately, is not open source.)
Significance
A unified theory of cortical function is proposed for guiding both neuroscience and artificial intelligence research. The theory offers an empirically testable framework for understanding how the brain accomplishes three key functions: (i) inference: perception is nonconvex optimization that combines sensory input with prior expectation; (ii) exploration: inference relies on neural response variability to explore different possible interpretations; (iii) prediction: inference includes making predictions over a hierarchy of timescales. These three functions are implemented in a recurrent and recursive neural network, providing a role for feedback connections in cortex, and controlled by state parameters hypothesized to correspond to neuromodulators and oscillatory activity.
Abstract
Most models of sensory processing in the brain have a feedforward architecture in which each stage comprises simple linear filtering operations and nonlinearities. Models of this form have been used to explain a wide range of neurophysiological and psychophysical data, and many recent successes in artificial intelligence (with deep convolutional neural nets) are based on this architecture. However, neocortex is not a feedforward architecture. This paper proposes a first step toward an alternative computational framework in which neural activity in each brain area depends on a combination of feedforward drive (bottom-up from the previous processing stage), feedback drive (top-down context from the next stage), and prior drive (expectation). The relative contributions of feedforward drive, feedback drive, and prior drive are controlled by a handful of state parameters, which I hypothesize correspond to neuromodulators and oscillatory activity. In some states, neural responses are dominated by the feedforward drive and the theory is identical to a conventional feedforward model, thereby preserving all of the desirable features of those models. In other states, the theory is a generative model that constructs a sensory representation from an abstract representation, like memory recall. In still other states, the theory combines prior expectation with sensory input, explores different possible perceptual interpretations of ambiguous sensory inputs, and predicts forward in time. The theory, therefore, offers an empirically testable framework for understanding how the cortex accomplishes inference, exploration, and prediction.
Introduction
Perception is an unconscious inference. Sensory stimuli are inherently ambiguous so there are multiple (often infinite) possible interpretations of a sensory stimulus (Fig. 1). People usually report a single interpretation, based on priors and expectations that have been learned through development and/or instantiated through evolution. For example, the image in Fig. 1A is unrecognizable if you have never seen it before. However, it is readily identifiable once you have been told that it is an image of a Dalmatian sniffing the ground near the base of a tree. Perception has been hypothesized, consequently, to be akin to Bayesian inference, which combines sensory input (the likelihood of a perceptual interpretation given the noisy and uncertain sensory input) with a prior or expectation.
Our brains explore alternative possible interpretations of a sensory stimulus, in an attempt to find an interpretation that best explains the sensory stimulus. This process of exploration happens unconsciously but can be revealed by multistable sensory stimuli (e.g., Fig. 1B), for which one’s percept changes over time. Other examples of bistable or multistable perceptual phenomena include binocular rivalry, motion-induced blindness, the Necker cube, and Rubin’s face/vase figure. Models of perceptual multistability posit that variability of neural activity contributes to the process of exploring different possible interpretations, and empirical results support the idea that perception is a form of probabilistic sampling from a statistical distribution of possible percepts. This noise-driven process of exploration is presumably always taking place. We experience a stable percept most of the time because there is a single interpretation that is best (a global minimum) with respect to the sensory input and the prior. However, in some cases, there are two or more interpretations that are roughly equally good (local minima) for bistable or multistable perceptual phenomena.
Prediction, along with inference and exploration, may be a third general principle of cortical function. Information processing in the brain is dynamic. Visual perception, for example, occurs in both space and time. Visual signals from the environment enter our eyes as a continuous stream of information, which the brain must process in an ongoing, dynamic way. How we perceive each stimulus depends on preceding stimuli and impacts our processing of subsequent stimuli. Most computational models of vision are, however, static; they deal with stimuli that are isolated in time or at best with instantaneous changes in a stimulus (e.g., motion velocity). Dynamic and predictive processing is needed to control behavior in sync with or in advance of changes in the environment. Without prediction, behavioral responses to environmental events will always be too late because of the lag or latency in sensory and motor processing. Prediction is a key component of theories of motor control and in explanations of how an organism discounts sensory input caused by its own behavior. Prediction has also been hypothesized to be essential in sensory and perceptual processing. ...Moreover, prediction might be critical for yet a fourth general principle of cortical function: learning.