Learning, Other Methods

Linear Homeostatic Plasticity

Formulation

\[ v(t + \Delta t) - v(t) = \lambda \frac{r^* - r}{r^*} \]

Where:

  • \(v\), an updatable parameter

  • \(r^*\), target spike rate

  • \(r\), observed spike rate

  • \(\lambda\), learning rate

  • \(t\), current simulation time

  • \(\Delta t\), duration of the simulation step

Description

This method is used for regulating the trainable parameters of a network trained with another plasticity method based on a target spiking rate. As originally provided, weight updates use a positive value for \(\lambda\) and delay updates use a negative value for \(\lambda\).

References

  1. DOI:10.1162/neco_a_01674