next up previous
Next: PROBABILITY ANALOGY Up: Probability Interpretation of Population Previous: Probability Interpretation of Population

INTRODUCTION

I define a ``population code'' or ``distributed representation'' to be a set of units for which the pattern of activity indicates the value of an underlying coded variable. This is in contrast to an ``analog representation'' in which the underlying variable is represented directly by the firing rates of one or more cells. The difference is illustrated in figure 1 (see [van Gelder 1991] for discussion). Perhaps the most common computational example of a distributed representation is the binary number system, in which a pattern of 1's and 0's represents an approximation to an analog value. In this paper I will consider a particular form of distributed representation in which each unit is active for only a small contiguous range of input values. A radial basis function representation with local basis functions has this form, as does the unary (but not the binary) number system. I will refer to such representations as having ``local receptive fields''.

  
Figure: a. ``Analog'', and b. ``Population'' coding of the number 3.

Population codes have been proposed as a model for information representation in certain regions of primate cortex [Georgopoulos et al. 1988,Andersen and Zipser 1988]. This paper seeks to provide the beginnings of a theoretical foundation for understanding such representations and their computational properties. In particular, I attempt to answer the following three questions:

  1. How can we compute functions of population coded variables, and are there supervised learning algorithms that can approximate desired functions?
  2. What is an appropriate training criterion for unsupervised learning and do the necessary algorithms exist?
  3. How can we interpret measurements made on biological population codes with reference to externally measurable variables?
I address these issues by first re-interpreting population codes with local receptive fields as representing estimates of the probability density of the input distribution. Using this interpretation, it becomes possible to apply the tools of probability theory to the analysis of population codes, and these tools yield simple yet powerful results that address the questions above.



next up previous
Next: PROBABILITY ANALOGY Up: Probability Interpretation of Population Previous: Probability Interpretation of Population



Terence D. Sanger
Mon Aug 21 18:36:58 EDT 1995