Redistribution of Synaptic Efficacy Supports Stable Pattern Learning in Neural Networks
MetadataShow full item record
Markram and Tsodyks, by showing that the elevated synaptic efficacy observed with single-pulse LTP measurements disappears with higher-frequency test pulses, have critically challenged the conventional assumption that LTP reflects a general gain increase. Redistribution of synaptic efficacy (RSE) is here seen as the local realization of a global design principle in a neural network for pattern coding. As is typical of many coding systems, the network learns by dynamically balancing a pattern-independent increase in strength against a pattern-specific increase in selectivity. This computation is implemented by a monotonic long-term memory process which has a bidirectional effect on the postsynaptic potential via functionally complementary signal components. These frequency-dependent and frequency-independent components realize the balance between specific and nonspecific functions at each synapse. This synaptic balance suggests a functional purpose for RSE which, by dynamically bounding total memory change, implements a distributed coding scheme which is stable with fast as well as slow learning. Although RSE would seem to make it impossible to code high-frequency input features, a network preprocessing step called complement coding symmetrizes the input representation, which allows the system to encode high-frequency as well as low-frequency features in an input pattern. A possible physical model interprets the two synaptic signal components in terms of ligand-gated and voltage-gated receptors, where learning converts channels from one type to another.