AUTHOR=Tsotsos John K. 

TITLE=Complexity Level Analysis Revisited: What Can 30 Years of Hindsight Tell Us about How the Brain Might Represent Visual Information?

JOURNAL=Frontiers in Psychology

VOLUME=Volume 8 - 2017

YEAR=2017

URL=https://www.frontiersin.org/journals/psychology/articles/10.3389/fpsyg.2017.01216

DOI=10.3389/fpsyg.2017.01216

ISSN=1664-1078

ABSTRACT=In a series of papers spanning 1987-2012, we examined the inherent computational difficulty of visual information processing using theoretical and empirical methods. The main goal of this activity had three components: to understand the deep nature of the computational problem of visual information processing; to discover how well the computational difficulty of vision matches to the fixed resources of biological seeing systems; and, to abstract from the matching exercise the key principles that lead to the observed characteristics of biological visual performance. The problem is clearly of interest to each of the communities studying vision, namely machine vision, neuroscience, psychology, cognitive science, artificial intelligence and robotics. This paper revisits those principles with the advantage that decades of hindsight can provide. It is clear, for example, that the current leading computational approaches to machine vision that employ deep learning have achieved their success in part due to conforming to the results of that analysis. It is also apparent that the problem of signal interference within a hierarchical network is significant and requires deeper examination. In order to deal with complexity issues and signal interference, we assert that the generality that human vision - and likely intelligence in general - exhibits is enabled by dynamic tuning of the brain's neural machinery based on task, environment and moment-by-moment requirements, and this is what attention accomplishes. And in order for this to occur, the underlying representations must be of a particular form: hierarchical, space and time limited, pyramidal, bidirectional, and dynamically tunable by task and context, which further requires each tunable representation to be semantically transparent.