mOja and AdaptiveLIFLateralInhibition

• added new mOja learning rule
• added new AdaptiveLIFLateralInhibition neuron type
• reformatted file by moving some common functions outside rules
• improved noise handling inside rules
• calculating memristor-based weights before first timestep
• added possibility to generate videos of weight heat maps
• expect to classify MNIST in an unsupervised way

New noise generation

Noise on parameters is now generated during the build process instead of at each timestep. This helps simulation speed and resolves the many numerical instabilities.

Input switching

• Added the possibility to specify two inputs: one for training, one for testing
• Made generate_sines() into a custom Nengo Process

v2.1

• Supports NengoDL Operator merging
• Supports new build format introduced in NengoDL 3.3.0
• Can pass memristor parameters to learning rule, including gain

Function learning using SrTiO3 memristors

Implementation of function learning using simulated SrTiO3 memristors. Runs on both Nengo Core and NengoDL.