Learning to Approximate Functions Using Nb-Doped SrTiO₃ Memristors.

Area of Science:

• Materials Science
• Computer Science
• Neuroscience

Background:

• Memristors are promising for neuromorphic computing due to their potential for efficient hardware implementation of artificial neurons and synapses.
• Interface-type memristors, specifically Nb-doped SrTiO3, are being investigated for their utility in neuromorphic hardware applications.

Purpose of the Study:

• To validate the use of Nb-doped SrTiO3 memristors as synaptic elements in a simulated neural network.
• To demonstrate a novel supervised learning algorithm for training memristor-based neural networks.

Main Methods:

• Nb-doped SrTiO3 memristors were arranged into differential synaptic pairs, with connection weights determined by normalized conductance differences.
• A simulated neural network utilized these memristor pairs and a supervised learning algorithm involving discrete voltage pulses.
• Noise was injected into the simulation to mimic the uncertainty of physical memristive devices and pulse impacts.

Main Results:

• The simulated neural network successfully learned to represent functions through the training process.
• Discrete updates based on local knowledge demonstrated robust learning performance despite simulated noise.
• The memristor-based network achieved universal function approximation capabilities.

Conclusions:

• Nb-doped SrTiO3 memristors are suitable for use as synaptic weight elements in spiking neural networks.
• This study presents one of the first models of a memristive spiking neural network capable of learning complex functions.
• These findings strongly suggest the potential of these memristors for future computing platforms.