Weight

Synonyms

Connection strength; Synaptic efficacy

Definition

In a neural network, connections between neurons typically have weights that indicate how strong the connection is. The neuron computes by forming a weighted sum of its input, i.e., the activation of each input neuron is multiplied by the corresponding connection weight. Adapting such weights is the most important way of learning in neural networks. Connection weights are loosely modeled after the synaptic efficacies in biological neurons, where they determine how large a positive or negative change in the membrane potential each input spike generates (see Biological Learning). In most models, all connection parameters are abstracted into a weight: attenuation or interaction of the potentials and connection delays are usually not taken into account. The weights are usually real-valued numbers ( − ∞ . . ∞), although in some algorithms, intended for VLSI implementation, the range and precision of these values can be restricted...