Plato's Camera by Churchland Paul M.;

Author:Churchland, Paul M.; [Churchland, Paul]
Language: eng
Format: epub
ISBN: 5652154
Publisher: MIT Press
Published: 2012-03-31T16:00:00+00:00

Figure 3.6

A simple recurrent network. See plate 14 or http://mitpress.mit.edu/platoimages to view this figure in color.

Their presence also means that the ways in which that lower rung responds to any sensory input, from still lower in the ladder, is no longer determined solely by the character of that sensory input and by the configuration of synaptic weights through which that input must pass. For the neuronal response at that now doubly connected rung will be partly a function of the prior activational state of the higher rung that projects those recurrent axons downward. That response will vary, for one and the same sensory input, as a function of the prior and ever-changing ‘cognitive context’ (supplied by activation in the higher population of neurons that project the relevant recurrent pathways) to which that sensory input happens to arrive. That is, the response will now be a function of the ever-changing dynamical ‘frame of mind’ of the perceiving network.

This structural/functional addition to our basic feed-forward neuronal machinery brings with it a wealth of novel and welcome behaviors. For starters, a recurrent network can have its synaptic weights configured (i.e., it can be trained) so that the trajectories within its activation space will tend, as a function of where they begin, to spiral quickly in toward one or another of a family of distinct attractor points. These points can perform the function, roughly, that the familiar prototype points served in the case of the feed-forward networks we have already examined. But in recurrent networks those special activation points can be reached after several swift cycles of recurrent activity, cycles that repeatedly exploit the background information embodied in the network’s synaptic connections. This often adds a further dimension of recognitional focus to the welcome phenomenon of vector completion that we have already noted, with approval, in the behavior of purely feed-forward networks. If the second rung of a recurrent network, given a slightly ambiguous or degraded sensory input, fails to zero in on a learned prototype on its first pass, by sheer feed-forward vector completion alone, it has the opportunity to ‘settle in’ on that prototype after several cycles of activation within the recurrent loop that feeds back to that rung. A prototype activation that failed to occur on the first hit from the input rung can nevertheless find a bull’s-eye realization after several cycles of recurrent assistance. A recurrent network can thus be one notch wiser than its purely feed-forward analogue, even where single points in activation space are concerned.

Well and good. But it is the recurrent network’s expanded capacity for representation that must command our primary attention. And the main enhancement they offer is their capacity to learn, not just a family of prototype points in the activation space of a given rung, but prototypical trajectories in the activation space of that rung. This allows them to represent prototypical sequences of events in the external world. It permits them to acquire a structured library of prototypical trajectories in

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.