Monolithic Optoelectronic Artificial Intelligence Neuromorphic Computing Hardware
- Lee, Yun-Jhu
- Advisor(s): Yoo, S. J. Ben
Abstract
Neuromorphic computing utilizes the computing method of the human brain to providesolutions to artificial intelligence. Developing Neuromorphic computing hardware is crucial to closely emulate how the human brain works. Spiking neural networks (SNN) is a kind of neuromorphic computing that provides a new computational paradigm capable of highly parallelized, real-time processing. Photonic spiking neural networks (PSNNs) potentially offer exceptionally low loss, low power, highly parallel, and high throughput compared to their electronic neuromorphic counterparts while maintaining their benefits in terms of event-driven computing capability. This thesis presents PSNN hardware utilizing monolithic silicon photonics (SiPh) process designs which consist of a Mach-Zehnder Interferometric (MZI) mesh incoherent network and event-driven laser optoelectronic spiking neurons. The Izhikevich-inspired optoelectronic spiking neuron consists of two photodetectors for excitatory and inhibitory optical spiking inputs, electrical transistors’ circuits providing spiking nonlinearity, and a laser for optical spiking outputs. There are two designs of optoelectronic spiking neurons demonstrated in the Section 2 and Section 3, respectively. One design focusing on the pursuit of energy efficiency (energy-efficient neuron) and another one focusing on the heterogeneity of neuron dynamics (heterogeneous neuron). Both event-driven neuron designs are inspired by the Izhikevich model incorporating both excitatory and inhibitory optical spiking inputs and producing optical spiking outputs accordingly. The neuron models are verified in Verilog- A and simulated the circuit-level operation of various cases with excitatory input and inhibitory input signals. The experimental results closely resemble the simulated results and demonstrate how the excitatory inputs trigger the optical spiking outputs while the inhibitory inputs suppress the outputs. Section 4 demonstrated a variation of the Random Backpropagation (RPB) learning algorithm on the MZI mesh synaptic interconnects and matched the performance of a standard linear regression on a simple classification task. The MZI mesh synaptic interconnects also connected with the heterogeneous neuron design in Section 3 to show the result of 89.3% accuracy on Iris image classification. The scenarios of multi-layer PSNN hardware packaging are also presented for future large-scale neuromorphic computing hardware. Section 5 presents the future direction for both energy-efficient neuron and heterogeneous neuron, which implemented with advanced transistor technology and show the benchmarking results. The nano energy-efficient neuron can achieve an estimated 21.09 fJ/spike input that can trigger the output from on-chip nanolasers running at a maximum of 10 Gspike/second in the neural network. Utilizing the simulated neuron model, the simulation results show 90% accuracy on unsupervised learning and 97% accuracy on a supervised modified FC neural network. The benchmark shows the PSNN design can achieve 50 TOP/J energy efficiency, which corresponds to 100x throughputs and 1000x energy-efficiency improvements compared to state-of-the-art electrical neuromorphic hardware such as Loihi and NeuroGrid. The nano heterogeneous neuron has power consumption of 1.18 pJ/spike output, and is predicted to achieve 36.84 fJ/spike output with a 7 nm CMOS platform (e.g. ASAP7) integrated with silicon photonics containing on-chip micron-scale lasers.