Paper on dynamics of superior colliculus & recurrent backprop.

[Lina Massone]

ftp-host: archive.cis.ohio-state.edu
ftp-file: massone.colliculus.ps.Z

The following paper has been placed in the connectionist archive. The 
paper has been accepted for publication in "Network".
(Note: the paper takes a LONG time to print because of the figures.)

	Local Dynamic Interactions in the Collicular Motor Map:
			A Neural Network Model

		Lina L.E. Massone and Tony Khoshaba

In this paper we explore the possibility that some of the dynamic properties
of the neural activity in the gaze-related motor map (located in the 
intermediate layers of the superior colliculus) might be mediated by local
interactions between movement-related neurons and fixation neurons.
More specifically, the goal of this research is to demonstrate, from a
computational standpoint, which classes of dynamic behaviors of the collicular
neurons can be obtained without the intervention of feedback signals, and
to hence begin exploring to what extent the gaze system needs feedback
in order to operate.
We modeled: (a) The collicular motor map as a dynamical system
realized with a recurrent neural network.  (b) The dynamics of the neural
activity in the map as the transients of that system towards an equilibrium
configuration that the network learned with a recurrent learning algorithm
(recurrent backpropagation.) The results of our simulations demonstrate:
(1) That the transients of the trained network are hill-flattening
patterns as observed by some experimenters in the burst-neuron layer of
the superior colliculus of rhesus monkeys. This result was obtained
despite the fact that the learning algorithm did not specify what the
network's transients should be.
(2) That the connections in the trained network are excitatory within the
fixation zone of the motor map and inhibitory elsewhere.
(3) That the results of the learning are robust in the face of changes
in the connectivity pattern and in the initialization of the weights, but that 
a local connectivity pattern favors the network's stability.
(4) That nonlinearity is required in order to obtain meaningful dynamic
behaviors.
(5) That the trained network is robust to abnormal stimulation patterns
such as noisy and multiple stimuli and that when multiple stimuli
are utilized the response of the network remains a stereotyped flattening
one.
The results of the learning point out the possibility that the 
dynamics of the burst-neuron layer of the superior colliculus might be
locally regulated rather than feedback-driven, and that the action of the
feedback be confined to the layer of the buildup neurons.
The results of the multiple-stimulation experiment support the hypothesis,
already put forward by one of the authors in a previous work (Massone
in press), that the averaging of the direction of movement following
double stimulation of the motor map (Robinson 1972) does not occur at the
level of the motor map.
This paper constitutes also  a study of the properties and responses
of recurrent backpropagation under various choices for the network's
and algorithm's parameters.