Optical Neural Network Accelerators

Deep neural networks (DNNs) are reshaping science and transforming people’s daily life. Their success in the past decade is due largely to the compute power and trainable datasets available. To tackle complex problems, the size of machine learning models scales exponentially, leading to a tremendous need for scaling up the compute power. However, improving the throughput of digital electronics has become difficult, not only because CMOS transistors cannot be made infinitely small (the slow-down of Moore’s Law), but also when the transistors get smaller, the power consumption per chip area goes up (kwon as the breakdown of Dennard scaling), due to charging wire capacitance while data movement. To overcome these bottlenecks, new computing paradigms leveraging different physics have to be developed. Our group focuses on optical neural networks, an emerging platform which unlocked the power of light with large optical bandwidth (e.g., 100s of THz) and nearly no loss in data movement. An example of a general-purposed optoelectronic tensor processor, homodyne optical neural network, is shown below, and its first realisation in a 3D architecture using arrays of VCSELs showing orders of magnitude improvement in throughput density and energy efficiency.

Coherent homodyne neural network architecture. a. forward propagating neural network. b. A deep neural network with N layers, each layer computes a vector-matrix multiplication. c. spatial-temporal multiplexed neural network, in each layer, the input vector is mapped to i time steps using a laser transmitter, whose beam is fanned out to j copies, each copy beat with a weight vector on a integrating photon-receiver that does the weighing accumulates. Ref: **Chen, et al, arXiv:2207.05329 (2022)**

Fig. 2 Experimental optical tensor processor. a. optimisation of compute density in a 3D photonic architecture. b. arrays of micron-scale lasers for high speed data transmission. Ref: **Chen, et al, arXiv:2207.05329 (2022)**

Low-Energy Edge Devices

Present-day sensor networks (e.g., with self-driving cars or smart home devices) reply on data acquisition at the edge and processing in the datacenter, where two-way communications lead to long latency in action execution and risk of data safety. Emerging decentralised network architectures using smart sensors with compute power (e.g., incorporating electronic micro-processors in edge devices) have allowed for in sensor processing with low latency; however, these electronic processors are power hungry and bulky in size, hampering the development of decentralised network technology. We approaches new innovations in the development of photonic edge processors for ultra-low energy, low-latency computing, as an example shown below.

Concept of photonic edge computing. The weights, trained at the datacenter, are broadcasted (using wavelength-multiplexing) to a photonic edge device for real-time inference. More information: **Sludds, et al, Science 378, 6617 (2022)**

Frequency Comb Interferometry (FCI) for Sensing, Imaging and Ranging

Laser frequency combs have revolutionised the field of optical metrology. As a frequency comb enabled technique, FCI is emerging to provide precision metrology with enhanced spectral resolution, time resolution, spectral coverage and sensitivity, etc. In a FCI system, one frequency comb (signal comb) interrogates a sample under study; a second comb of slightly different repetition frequency (e.g., generated by a Doppler scanner in a Michelson interferometer or a separate laser cavity of slightly different length) serves as a sampling oscilloscope in time domain or a spectrometer in the frequency domain to readout the signal comb. FCI allows broadband spectroscopic measurements with spectral coverages up to octave spanning, resolution exceeding to the comb spacing and data acquisition speed freely set by the repetition rate difference of the two combs. FCI has been widely exploited for laboratory research, our group is dedicated in exploring new applications of FCI for real-life applications.