What is the goal of sensory coding?

  • Authors:

  • David J. Field

  • Affiliations:

  • -

  • Venue:
  • Neural Computation
  • Year:
  • 1994

Quantified Score

Hi-index 0.00

Visualization

Abstract

A number of recent attempts have been made to describe earlysensory coding in terms of a general information processingstrategy. In this paper, two strategies are contrasted. Bothstrategies take advantage of the redundancy in the environment toproduce more effective representations. The first is described as a"compact" coding scheme. A compact code performs a transform thatallows the input to be represented with a reduced number of vectors(cells) with minimal RMS error. This approach has recently becomepopular in the neural network literature and is related to aprocess called Principal Components Analysis (PCA). A number ofrecent papers have suggested that the optimal compact code forrepresenting natural scenes will have units with receptive fieldprofiles much like those found in the retina and primary visualcortex. However, in this paper, it is proposed that compact codingschemes are insufficient to account for the receptive fieldproperties of cells in the mammalian visual pathway. In contrast,it is proposed that the visual system is near to optimal inrepresenting natural scenes only if optimality is defined in termsof "sparse distributed" coding. In a sparse distributed code, allcells in the code have an equal response probability across theclass of images but have a low response probability for any singleimage. In such a code, the dimensionality is not reduced. Rather,the redundancy of the input is transformed into the redundancy ofthe firing pattern of cells. It is proposed that the signature fora sparse code is found in the fourth moment of the responsedistribution (i.e., the kurtosis). In measurements with 55calibrated natural scenes, the kurtosis was found to peak when thebandwidths of the visual code matched those of cells in themammalian visual cortex. Codes resembling "wavelet transforms" areproposed to be effective because the response histograms of suchcodes are sparse (i.e., show high kurtosis) when presented withnatural scenes. It is proposed that the structure of the image thatallows sparse coding is found in the phase spectrum of the image.It is suggested that natural scenes, to a first approximation, canbe considered as a sum of self-similar local functions (the inverseof a wavelet). Possible reasons for why sensory systems wouldevolve toward sparse coding are presented.