Synaptic plasticity

Changes in synaptic efficacy over time. Activity-dependent plasticity can support learning and development, but also requires constraints to prevent unbounded growth or competition in simplified models.

A decaying record of recent activity that marks a state or connection as eligible for modification by a later teaching signal. An accumulating value-learning trace is $e_t=\gamma \lambda e_{t-1}+x(s_t)$, giving $\Delta w=\alpha {\delta }_t{e}_t$. Other traces record action-policy gradients or pre/post synaptic events. The update and episode-boundary convention must be specified; an eligibility trace is not automatically identical to the visual activity trace used for trace learning.

A persistent increase in synaptic efficacy induced by an appropriate pattern of activity. Its presence is a possible storage mechanism; a memory model must still explain which inputs are modified, the learning conditions and how subsequent activity retrieves the stored information.

Short-lived changes in synaptic efficacy dependent on recent activity, including facilitation and depression. Such changes can alter a network's response to a later input even when its immediately recorded activity does not reveal the retained information.

Constraints or interactions make the strengthening of some inputs limit others. In developmental models, competition together with input correlations can segregate ON and OFF subregions or inputs from the two eyes; correlations alone do not specify the full biological developmental mechanism.

A learning rule can constrain the sum or norm of incoming synaptic weights to prevent unbounded Hebbian learning. For ${\sum }_i{w}_{ai}=S$, subtracting the mean unconstrained update from each update preserves the sum. Bounds and renormalization may also be needed to keep excitatory weights nonnegative. Normalization of learning weights is distinct from divisive gain control in instantaneous neural responses.