Input Prediction and Autonomous Movement Analysis in Recurrent Circuits
of Spiking Neurons
R. Legenstein, H. Markram, and W. Maass
Temporal integration of information and prediction of future sensory inputs are
assumed to be important computational tasks of generic cortical
microcircuits. However it has remained open how cortical microcircuits could
possibly achieve this, especially since they consist in contrast to most
neural network models of neurons and synapses with heterogeneous dynamic
responses. However it turns out that the diversity of computational units
increases the capability of microcircuit models for temporal integration.
Furthermore the prediction of future input may be rather easy for such
circuits since it suffices to train the readouts from such microcircuits. In
this article we show that very simple readouts from a generic recurrently
connected circuit of integrate-and-fire neurons with diverse dynamic synapses
can be trained in an unsupervised manner to predict movements of different
objects, that move within an unlimited number of combinations of speed,
angle, and offset over a simulated sensor field. The autonomously trained
microcircuit model is also able to compute the direction of motion, which is
a computationally difficult problem ("aperture problem") since it requires
disambiguation of local sensor readings through the context of other sensor
readings at the current and preceding moments. Furthermore the same circuit
can be trained simultaneously in a supervised manner to also report the shape
and velocity of the moving object. Finally it is shown that the trained
neural circuit supports novelty detection and the generation of "imagined
movements". Altogether the results of this article suggest that it is not
necessary to construct specific and biologically unrealistic neural circuit
models for specific sensory processing tasks, since "found" generic cortical
microcircuit models in combination with very simple perceptron-like readouts
can easily be trained to solve such computational tasks.
Reference: R. Legenstein, H. Markram, and W. Maass.
Input prediction and autonomous movement analysis in recurrent circuits of
Reviews in the Neurosciences (Special Issue on Neuroinformatics of Neural
and Artificial Computation), 14(1-2):5-19, 2003.