Study On Programmable Time-Wavelength Interleaving Optical Neural Network

The era of Big Data has provided fertile soil for the development of the Fourth Industrial Revolution powered by artificial intelligence(AI).As the driving force of AI industrial,artificial neural network(ANN)algorithm related machine learning(ML)technology is leading dramatic changes in many spheres such as vision,voice and natural language,which enables advanced progresses with each passing day.However,it is become overwhelmed for conventional electronic computing hardware to power and adapt the continuing-developed ANN algorithm,and the need for advanced parallel computing architecture becomes to be urgent accordingly.Optics presents its unique potential for next generation high performance computing due to its intrinsic parallelism,ultrahigh bandwidth and powerful reconfigurability.In addition,with the fast development of silicon photonics technology,integrated optical computing has become an outstanding candidate to break the performance bottleneck of electronic computing chip and continue the Moore's Law.In this dissertation,a time-wavelength constraint plane is established by broadband modulation of a wavelength division multiplexing(WDM)light source,and the interleaved manipulation of time and wavelength is accordingly carried out.Combined with time division multiplexing(TDM)technique,an advanced optical matrix multiplication(OM²U)is proposed.We demonstrate the availability of this architecture in performing matrix operation by both simulation and experiment with computation speed up to 1.18 x 10¹¹ MACs per second.We also explore an integrated scheme for OM²U with microring modulator(MRM)array and balanced photodetector(BPD),and an optical neural network(ONN)for classification task is simulated by reusing single OM²U with 90%accuracy and 1.25 x 10⁹ MACs per second computation speed.The proposed architecture shows fine scalability and flexibility with relative simple hardware structure by utilizing multi-dimensional optical signal processing and multiplexing techniques,which is capable of arbitrary layers high-accuracy ANN algorithm acceleration.Besides,universal optical components are explored in building OM²U,showing promising potential for massive fabrications and applications.