All-Chalcogenide Programmable All-Optical Deep Neural Networks

Abstract

Deeplearning algorithms are revolutionising many aspects of modern life. Typically, they are implemented in CMOS-based hardware with severely limited memory access times and inefficient data-routing. All-optical neural networks without any electro-optic conversions could alleviate these shortcomings. However, an all-optical nonlinear activation function, which is a vital building block for optical neural networks, needs to be developed efficiently on-chip. Here, we introduce and demonstrate both optical synapse weighting and all-optical nonlinear thresholding using two different effects in one single chalcogenide material. We show how the structural phase transitions in a wide-bandgap phase-change material enables storing the neural network weights via non-volatile photonic memory, whilst resonant bond destabilisation is used as a nonlinear activation threshold without changing the material. These two different transitions within chalcogenides enable programmable neural networks with near-zero static power consumption once trained, in addition to picosecond delays performing inference tasks not limited by wire charging that limit electrical circuits; for instance, we show that nanosecond-order weight programming and near-instantaneous weight updates enable accurate inference tasks within 20 picoseconds in a 3-layer all-optical neural network. Optical neural networks that bypass electro-optic conversion altogether hold promise for network-edge machine learning applications where decision-making in real-time are critical, such as for autonomous vehicles or navigation systems such as signal pre-processing of LIDAR systems.

On-Chip Integrated Photonic Devices Based on Phase Change Materials

“…As for the visible region, Sb 2 S 3 still retrains near-zero absorption as opposed to Sb 2 Se 3 , which shows a sharper rise in the absorption coefficient [40]. The property of low loss attracted renewed interest in these materials [41][42][43][44][45][46], for both on-chip and freespace applications. On the flip side, Sb 2 S 3 reportedly also undergoes larger size variation when switching from amorphous to the crystalline phase.…”

“…The neural network, made using of Mach-Zehnder interferometers (MZIs) with GSST on both arms, was trained and tested using MNIST [135] handwritten digits and was able to achieve a high level of accuracy. (c,d) reproduced under creative commons license [23]; (e,f) reproduced under creative commons license from article in arXiv [46]).…”

“…On the speed and energy efficiency front, the 'read' and 'write' actions take 500 ps and 1.5 ns for 1 pJ and 4 pJ, respectively, using pulses with a duration of 200 ps. This system was later further (c,d) reproduced under creative commons license [23]; (e,f) reproduced under creative commons license from article in arXiv [46]).…”

“…Scaling the neural network further, [25] recently presented a synapse, graphical processing unit, and a tensor processing unit that perform in-memory calculations through the network of coupler-based latch circuit using Si 3 N 4 with Sb 2 S 3 , as well as GST as the PCM [46]. The network (shown in Figure 8e,f) performs at the maximum data rate of 100 GHz and consumes 0.1 uJ of energy.…”

“…Nonlinear variations in refractive index are sensed with TM polarized light. ((a,b) reproduced under creative commons license [95];(c,d) reproduced under creative commons license[23]; (e,f) reproduced under creative commons license from article in arXiv[46]).…”