As opposed to conventional computing systems based on the von Neumann architecture (in which the central computing unit is separate from the main memory), the human brain contains a vast number of neurons and synapses that act both as the computing and memory units. This unique structure means that we can deal with emotions, learn and think using hardly any energy. These actions are nigh-on impossible, or at least far less efficient, with traditional computers.

A neuron works by generating action potentials (spiking with a fire time of tpre) that propagate along the axon. These action potentials are transmitted through a junction to the next neuron (known as the post-neuron) which generates postsynaptic potentials (spiking with a fire time of tpost). This junction is the synapse and has a “synaptic weight” (w), which determines the communication strength between the two pre- and post-neurons.

Changes in synaptic weight (Δw) are known as “synaptic plasticity”, which depends on neural activities. For example, Δw=f(tpost-tpre) is thought to be the main mechanism by which human and animal brains learn and memorize things (the Hebbian learning rule).

Mimicking the neural system at the physical level

Neuromorphic, or brain-inspired, computing aims to mimic the neural system at the physical level of neurons and synapses and will rely on neuronal-like networks rather than series of binary 1s and 0s. It will be able to more easily handle the vast data sets currently being generated around the world (big data) and support emerging technologies like artificial intelligence and the internet of things (IoT), thanks to being massively parallel. To make this new generation of machines, however, researchers need to develop suitably plastic synapse-like devices – not least because synapses far outnumber neurons by several orders of magnitude in the human brain.

Electronic devices based on electrically-induced resistive changes in phase-change chalcogenides, metal–insulator–metal structures, ferroelectric materials and nanomaterial-based field-effect transistors show promise here. Photonic synapses based on microfibres and optoelectronic synapses made using carbon nanotubes are also good contenders. The problem is that such devices are difficult to integrate into chips and are limited in speed. They are not truly all-optical either because they still need to be excited with electrical signals to operate.

Synaptic weighting and processing

A team led by Harish Bhaskaran has now made a photonic synapse that uses the phase-change chalcogenide Ge2Sb2Te5 combined with integrated silicon nitride waveguides that transmits optical signals for both synaptic weighting and processing. “We can modulate the optical transmission through the waveguide by the number of pulses sent through it,” explains Bhaskaran. “However, we weight real synapses according to the time delay between pre-neuronal and post-neuronal spikes and propose a technique to convert such pulse number dependency to a time-delay dependency which, in effect, can mimic synaptic plasticity.”

The researchers say that they employed an optical circulator for connecting the output of the synapse and the post-neuron, and for applying optical pulses to modify the synaptic weight (see image). They can measure low-energy optical transmission from a pre-neuron to a post-neuron with the transmission level depending on the synaptic weight.

Modifying the multi-photonic synaptic weighting

“The Ge2Sb2Te5 can be used to modulate light (a property that allowed for rewritable CDs and DVDs),” explains team member and first author of the study Zengguang Cheng. “The synapse resembles neural synapses at the physical level and can achieve synaptic plasticity like that seen in Hebbian learning or the spike timing-dependent plasticity (STDP) rule.

“As mentioned, we modified the multi-photonic synaptic weighting by simply varying the number of optical pulses sent down the waveguide,” he says. “By optimizing the optical pulses and the device structures employed, we believe that we will be able to reach a continuously variable synaptic plasticity resembling that of real biological synapses. Our work is in fact the first step to show that it is, in principle, possible to do this.

“Optical computing is a pretty hot topic since it is the last bastion of optics (taking fibre optics all the way to the computer chip,” Bhaskaran tells nanotechweb.org. “It has the same advantages at the chip scale as it does on a large scale – that is, ultrafast operation speeds, unlimited bandwidth and no electrical power losses from parasitic components. Obviously, it is a very new field, so we still need to understand what the eventual limitations might be.

Still early days

“The photonic synapse is the first step for photonic neuromorphic computing and we can imagine plenty of applications – in artificial intelligence, big data and IoT. However, we must emphasize that this is still early days and we need many more innovations before being able to optically compute with on-chip neurons.”

The team, reporting its work in Science Advances DOI: 10.1126/sciadv.1700160, says that it will now be trying to make a photonic firing neuron and a complete architecture for photonic neuromorphic computing. “In the long term, our research collaboration, which includes physicists and engineers from the University of Münster in Germany and the University of Exeter in the UK, would like to help build entire computing platforms using such devices on a chip,” says Bhaskaran.