Design and Evaluation of Photonic Neural Network for Hardware-Accelerated Artificial Intelligence Applications

Nearly two-thirds of the global population now have internet access. As a result, annual global Internet traffic has increased exponentially to 4.8 Zettabytes (ZB) by 2022. The rise of users has driven the growth of data centers and smart devices for Internet-of-Things (IoT), expected to take a share of 23% of the global energy footprint by 2030. This problem is going mutually with the von-Neumann computing architecture bottleneck. The saturation of clock frequency in the central processing unit (CPU) has flattened out the current computing machine efficiency, which cannot sustain future data processing and generation demand. Eventually, the digital brick wall leads to a computing gap where a new computing technology is necessary.

On the other hand, the brain is a powerful natural computer within humans, which has a large scale and an efficient algorithm beyond the human-built computer. The brain has the capability to perform 10¹³ - 10¹⁶ operations per second with a power of only 15-25 W, which is eight orders of magnitude more efficient than the current digital electronic supercomputer. The parallel processing of the brain's neural network is the core of high-speed computation, which has been studied to develop artificial intelligence (AI). The success of AI, which has empowered decision-making machines, has pushed the investigation of brain-inspired computers at the physical level or a so-called neuromorphic computing.

Photonics, with its inherent parallelism and proven ability to offer low latency and fast signal processing in the GHz range, is a promising platform for neuromorphic computing. This potential of photonics for neuromorphic computing offers hope for the future of technology, with the possibility of developing faster and more efficient computing systems.

The neural network is based on two essential ingredients: the linear weighted summation, where all the neural inputs are weighted and summed, followed by the nonlinear activation function for the decision boundary. One approach to performing analog neuromorphic processing is to employ it in the optical domain, where the equivalent neural network functions have to be identified. Towards this direction, this thesis plays an important role in finding the key ingredients in the photonics realm based on the frequency coding of the synaptic signals.

The first part presents a synaptic receptor that functionally integrates the weighting and summation functions. The matrix multiplication based on the frequency coding of synaptic signals can be performed by employing the wide bandwidth of a semiconductor optical amplifier (SOA) and an electroabsorption modulator (EAM). It will be shown that two spike trains can be simultaneously processed with opposite polarities and detected as the weighted sum. The result proves that the synaptic receptor can carry multiple colors of synapses and can potentially scale the optical neural network (ONN).

Afterwards, this thesis investigates the photonic activation function, focusing on the rectified linear unit (ReLU) function. The ReLU unit is performed by optical frequency coding of signals via a linear transmission of an optical filter and a subsequent photodetector with its noise for clipping the signal. A digital neural network (DNN) based on the Iris flower classification case is used as a benchmark. It will be demonstrated that a small penalty of 3% in accuracy is achieved when transferring the classification challenge from the digital to the optical domain. This renders that the scheme can be efficiently employed in the ONN.

Next, the constituent ingredients are combined to perform the photonic neuron. Multiple neural sub-circuits can be collapsed over the photonic ReLU. The demonstration reveals that the weighting functionality can be integrated over the same ReLU unit, which motivates the simplification and cost-optimization of ONN scaling. A low penalty of 1% can be achieved in terms of accuracy compared to DNN.

Nevertheless, optoelectronic conversion is the main bottleneck between the linear and nonlinear operations in the photonic neuron. A translucent concatenation scheme is discussed afterwards, where the photocurrent of the weighted sum directly drives the subsequent FM translator in the activation circuit without the need for electronic amplification. This investigation promises an energy-efficient interface that can be scaled to a smaller footprint.

Eventually, the functional complete neuron in neuromorphic processing is presented in a space-switch architecture. Especially, the space-switch architecture can be made reconfigurable as a so-called photonic processing unit (ΦPU) to employ other signal processing functions, taking advantage of re-using the optical elements. It will be confirmed that ΦPU has a penalty-free operation. Furthermore, the neuromorphic processing in the space-switch is compared with the one in wavelength-routed architecture by employing the translucent concatenation scheme. A penalty <2% can be achieved for multiple optical/digital sub-circuits implementation scenarios.

The novelties presented in the thesis contribute to the neuromorphic photonics field and the possibility of scaling the ONN in a smaller device footprint.