Event-driven adaptive optical neural network

Abstract

We present an adaptive optical neural network based on a large-scale event-driven architecture. In addition to changing the synaptic weights (synaptic plasticity), the optical neural network's structure can also be reconfigured enabling various functionalities (structural plasticity). Key building blocks are wavelength-addressable artificial neurons with embedded phase-change materials that implement nonlinear activation functions and nonvolatile memory. Using multimode focusing, the activation function features both excitatory and inhibitory responses and shows a reversible switching contrast of 3.2 decibels. We train the neural network to distinguish between English and German text samples via an evolutionary algorithm. We investigate both the synaptic and structural plasticity during the training process. On the basis of this concept, we realize a large-scale network consisting of 736 subnetworks with 16 phase-change material neurons each. Overall, 8398 neurons are functional, highlighting the scalability of the photonic architecture.

Fig. 1.. Computing neural networks on the event-driven architecture.

(A) The system consists of waveguide-coupled PCM cells that emulate the functionality of artificial neurons. We address every PCM neuron on-chip with a distinct wavelength. To plastically connect the ONN, we encode the synapses in the event queue which is driving the system. This way, we can adapt and change the structure of the whole NN without having to modify the photonic circuit itself. (B) We operate the system sequentially, executing one event after the other. For each event, we send a set of input pulses (the synaptic configuration information) to the PCM neurons simultaneously. The wavelengths of the pulses correspond to the wavelength addresses of the presynaptic neurons and the pulse powers to the synaptic weights. The pulse powers are multiplied with the respective presynaptic PCM neuron activation and are afterward summed by a photodetector (PD). To conclude the event, a high-power write pulse is sent to the ONN, with a wavelength corresponding to the address of the postsynaptic neuron and a pulse power proportional to the output voltage of the PD. In this way, the state of the postsynaptic neuron is set. (C) Correspondence between neural nets and their optical implementation. Each physical operation performed by the ONN (green, lower row) can be directly mapped to the computation steps in an artificial NN (orange, upper row).

Fig. 2.. Photonic circuit implementing a subset of 16 artificial neurons.

(A) Input pulses are sent to the PCM neurons via a common bus waveguide. Microring resonators select a pulse of a certain wavelength from the bus waveguide, let the pulse interact with the PCM cell and couple it back to the bus waveguide in reverse direction. (B) The scanning electron microscope shows a single optical neuron. It consists of a critically coupled ring resonator for wavelength addressing, a PCM cell on top of an MMI focusing structure and a compact Bragg mirror to reflect the light back to the bus waveguide. The insets show a zoom of the Bragg mirror and the PCM cell. (C) The transmission spectrum of the photonic circuit in (A) shows 16 clearly distinguishable resonance peaks, each one corresponding to a different optical PCM neuron. (D) We obtain both excitatory and inhibitory behavior of the neuron by defining a partially amorphous state as the ground state in our neural network and switch the cell depending on the input pulse power. The bar length shows the SD of the transmission for each PCM state when randomly switching between the states.

Fig. 3.. Evolutionary learning with an adaptive ONN.

(A) We use an evolutionary algorithm to learn an event queue for distinguishing between German and English text samples. First, we randomly initialize a set of neural networks. Next, we perform a crossover between the neural networks, combining synapses from one neural network with those of another one. Then, the parent event queues are randomly mutated, and a random queue is added to the population. Last, we choose a subset of all event queues based on their performance. If the performance is sufficient, then we stop the learning algorithm; otherwise, we repeat the mating process on the selected queues. (B) We use an evolutionary algorithm to train the ONN on-chip. Apart from changes in the synaptic weights, the structure of the network is also modified. (C) Already after the second evolution, the network can clearly distinguish between the text samples.

Fig. 4.. Scalability of the adaptive ONN.

(A) We characterize the performance of an ONN consisting of 11,776 different PCM neurons arranged in a 32 × 23 array of separate devices. Each device contains 16 neurons. (B) For each PCM neuron, we measure the overall transmission in the as-sputtered amorphous and fully crystalline state. We obtain a switching contrast of 16 dB between the amorphous (−18.25 ± 0.95 dB) and crystalline state (−33.99 ± 2.54 dB). Some GST cells are not deposited at all, resulting in the blue peak of around 500 nominally “crystallized” neurons at −20 dB due to a saturated detector. (C) Typically, the neurons in a subnetwork are either mostly functional and show a contrast between the amorphous and crystalline state of at least 10 dB, or completely broken due to a fabrication error in the coupling or routing. In total, 8398 neurons are functional.