A Robust Learning Rule for Memristor-based Synapses Competitive with Supervised Learning in Standard Spiking Neural Networks

Paper

TBA

Abstract

Memristive devices are a class of circuit elementsthat shows great promise as future building block for brain-inspired computing. One influential view in theoretical neuro-science sees the brain as a function-computing device: giveninput signals, the brain applies a function in order to generatenew internal states and motor outputs. Therefore, being able toapproximate functions is a fundamental axiom to build upon forfuture brain research and to derive more efficient computationalmachines. In this work we apply a novel supervised learning algo-rithm - based on controlling niobium-doped strontium titanatememristive synapses - to learning non-trivial multidimensionalfunctions. By implementing our method into the spiking neuralnetwork simulator Nengo, we show that we are able to atleast match the performance of the standard Prescribed ErrorSensitivity learning rule, which is similar to the delta rule inclassical neural networks.

Running the code

1. Clone this repository for the library code and add it to your PYTHONPATH
2. Run the experiments:
   • memristor_evolution_test.py plots the power-law of the memristors
   • learn_multidimensional_functions.py runs mPES, and PES learning alongside NEF by using the simulated memristors in the memristor_nengo library.
     • Learn the product of 2-D input components (x1 * x2).
     • Learn the combined product (x1 * x2 + x3 * x4).
     • Learn the separate 3-D products [x1 * x2, x1 * x3, x2 * x3].
     • Learn the 2-D circular convolution [x1, x2] x [x3, x4].
     • Learn the 3-D circular convolution [x1, x2, x3] x [x4, x5, x6].