Physical deep learning with biologically inspired training method: gradient-free approach for physical hardware.

Ever-growing demand for artificial intelligence has motivated research on unconventional computation based on physical devices. While such computation devices mimic brain-inspired analog information processing, the learning procedures still rely on methods optimized for digital processing such as backpropagation, which is not suitable for physical implementation. Here, we present physical deep learning by extending a biologically inspired training algorithm called direct feedback alignment. Unlike the original algorithm, the proposed method is based on random projection with alternative nonlinear activation. Thus, we can train a physical neural network without knowledge about the physical system and its gradient. In addition, we can emulate the computation for this training on scalable physical hardware. We demonstrate the proof-of-concept using an optoelectronic recurrent neural network called deep reservoir computer. We confirmed the potential for accelerated computation with competitive performance on benchmarks. Our results provide practical solutions for the training and acceleration of neuromorphic computation.

Cross-gain modulation-based photonic reservoir computing using low-power-consumption membrane SOA on Si.

We demonstrate photonic reservoir computing (RC) utilizing cross-gain modulation (XGM) in a membrane semiconductor optical amplifier (SOA) on a Si platform. The membrane SOA's features of small active volume and strong optical confinement enable low-power nonlinear operation of the reservoir, with 101-mW-scale power consumption and 102-µW-scale optical input power. The power consumption is about an order of magnitude lower than that of conventional SOAs that exhibit saturable nonlinearity. The XGM-based reservoir is configured by injecting a delayed feedback signal into the SOA from a direction opposite to the input signal. This configuration provides robust operation of the feedback circuit because of the phase insensitivity and the elimination of loop oscillation risk. The RC performance is evaluated via the information processing capacity (IPC) and a nonlinear benchmark task. It is revealed that the XGM-based reservoir performs strong nonlinear transformation of input time-series signals. The series of results consistently show that the membrane SOA performs RC-applicable nonlinear operations through XGM at a low power scale.

Neural Schrödinger Equation: Physical Law as Deep Neural Network.

We show a new family of neural networks based on the Schrödinger equation (SE-NET). In this analogy, the trainable weights of the neural networks correspond to the physical quantities of the Schrödinger equation. These physical quantities can be trained using the complex-valued adjoint method. Since the propagation of the SE-NET can be described by the evolution of physical systems, its outputs can be computed by using a physical solver. The trained network is transferable to actual optical systems. As a demonstration, we implemented the SE-NET with the Crank-Nicolson finite difference method on Pytorch. From the results of numerical simulations, we found that the performance of the SE-NET becomes better when the SE-NET becomes wider and deeper. However, the training of the SE-NET was unstable due to gradient explosions when SE-NET becomes deeper. Therefore, we also introduced phase-only training, which only updates the phase of the potential field (refractive index) in the Schrödinger equation. This enables stable training even for the deep SE-NET model because the unitarity of the system is kept under the training. In addition, the SE-NET enables a joint optimization of physical structures and digital neural networks. As a demonstration, we performed a numerical demonstration of end-to-end machine learning (ML) with an optical frontend toward a compact spectrometer. Our results extend the application field of ML to hybrid physical-digital optimizations.

Compound-eye metasurface optics enabling a high-sensitivity, ultra-thin polarization camera.

Polarization imaging is key for various applications ranging from biology to machine vision because it can capture valuable optical information about imaged environments, which is usually absent in intensity and spectral content. Conventional polarization cameras rely on a traditional single-eye imaging system with rotating polarizers, cascaded optics, or micropolarizer-patterned image sensors. These cameras, however, have two common issues. The first is low sensitivity resulting from the limited light utilization efficiency of absorptive polarizers or cascaded optics. The other is the difficulty in device miniaturization due to the fact that these devices require at least an optical-path length equivalent to the lens's focal length. Here, we propose a polarization imaging system based on compound-eye metasurface optics and show how it enables the creation of a high-sensitivity, ultra-thin polarization camera. Our imaging system is composed of a typical image sensor and single metasurface layer for forming a vast number of images while sorting the polarization bases. Since this system is based on a filter-free, computational imaging scheme while dramatically reducing the optical-path length required for imaging, it overcomes both efficiency and size limitations of conventional polarization cameras. As a proof of concept, we demonstrated that our system improves the amount of detected light by a factor of ∼2, while reducing device thickness to ∼1/10 that of the most prevalent polarization cameras. Such a sensitive, compact, and passive device could pave the way toward the widespread adoption of polarization imaging in applications in which available light is limited and strict size constraints exist.