Summary: Neurocomputing 5860 (2004) 327335
Control of network activity through neuronal
Christian D. Swinehart , Kris Bouchard, Peretz Partensky,
Volen Center for Complex Systems and Department of Biology, Brandeis University,
MS 13, Waltham, MA 02454, USA
Neural network learning is typically treated as the problem of setting synaptic connection
strengths to better perform a task. This requires supervision. However, anatomical data suggest
that direct synaptic modi˙cation by such a supervisor circuit would be unfeasible. We investigate
supervision at the level of neurons rather than synapses. By modulating the response properties
of cells in the network, this form of supervised learning is able to successfully train a network to
perform a function approximation task. We examine the nature of this modulation, and consider
its implications for supervised learning.
c 2004 Elsevier B.V. All rights reserved.
Keywords: Neural networks; Learning; Modulation; Supervision
Neural networks that perform a speci˙c task must be developed through a learning