Photonic Neuromorphic Engineering

Over the recent years, we have investigated the properties of quantum wells (QWs) at the LaAlO3/SrTiO3 interface, including 2D superconductivity, Rashba spin-orbit fields and lattice vibrational modes [see References 1-3 below]. More recently we have uncovered persistent photoconductance (PPC), whereby the system changes its conductance in a plastic way, retaining memory from its past history, as in the case of memristors, but using light instead of electric pulses. Our most recent discovery [4] is that light pulses can be used to replicate spike timing-dependent plasticity (STDP). STDP was proposed to emulate time causality of electro-chemical signals in biological neurons: pre-synaptic neurons spiking after post-synaptic neurons are “anti-causal” and learning is weakened; pre-synaptic neurons spiking before post-synaptic neurons are causal, reinforcing learning. STDP enables unsupervised learning, without need of labelling training data.

Our discovery is particularly relevant, as it extends the STDP concept beyond electrical stimuli to the realm optical stimuli, opening up whole new perspectives on neuromorphic engineering and in artificial vision.  We develop algorithms that will help to understand how artificial networks can be designed to learn from visual inputs, with the ultimate objective of building a first design that may learn from simple visual patterns.

figure gervasi herranz
(Left) We can emulate STDP of biological neurons by illuminating the QWs with pairs of short-(violet)/long- (red) wavelength pulses of visible light. (Right) The reinforcement/depression of synaptic strength –represented by the photoconductance of the QW– might be exploited in artificial vision networks to replicate spatial and navigation maps built in biological brains.

[1] Pesquera et al., Physical Review Letters 2014.

[2] Herranz et al., Nature Communications 2015.

[3] Gazquez et al., Physical Review Letters 2017.

[4] Y. Chen et al., submitted.