On-chip diffractive optical neural networks (DONNs) bring the advantages of parallel processing and low energy consumption. However, an accurate representation of the optical field’s evolution in the structure cannot be provided using the previous diffraction-based analysis method. Moreover, the loss caused by the open boundaries poses challenges to applications. A multimode DONN architecture based on a more precise eigenmode analysis method is proposed. We have constructed a universal library of input, output, and metaline structures utilizing this method, and realized a multimode DONN composed of the structures from the library. On the designed multimode DONNs with only one layer of the metaline, the classification task of an Iris plants dataset is verified with an accuracy of 90% on the blind test dataset, and the performance of the one-bit binary adder task is also validated. Compared to the previous architectures, the multimode DONN exhibits a more compact design and higher energy efficiency.

As an optical processor, a diffractive deep neural network (D²NN) utilizes engineered diffractive surfaces designed through machine learning to perform all-optical information processing, completing its tasks at the speed of light propagation through thin optical layers. With sufficient degrees of freedom, D²NNs can perform arbitrary complex-valued linear transformations using spatially coherent light. Similarly, D²NNs can also perform arbitrary linear intensity transformations with spatially incoherent illumination; however, under spatially incoherent light, these transformations are nonnegative, acting on diffraction-limited optical intensity patterns at the input field of view. Here, we expand the use of spatially incoherent D²NNs to complex-valued information processing for executing arbitrary complex-valued linear transformations using spatially incoherent light. Through simulations, we show that as the number of optimized diffractive features increases beyond a threshold dictated by the multiplication of the input and output space-bandwidth products, a spatially incoherent diffractive visual processor can approximate any complex-valued linear transformation and be used for all-optical image encryption using incoherent illumination. The findings are important for the all-optical processing of information under natural light using various forms of diffractive surface-based optical processors.

Object identification and three-dimensional reconstruction techniques are always attractive research interests in machine vision, virtual reality, augmented reality, and biomedical engineering. Optical computing metasurface, as a two-dimensional artificial design component, has displayed the supernormal character of controlling phase, amplitude, polarization, and frequency distributions of the light beam, capable of performing mathematical operations on the input light field. Here, we propose and demonstrate an all-optical object identification technique based on optical computing metasurface, and apply it to 3D reconstruction. Unlike traditional mechanisms, this scheme reduces memory consumption in the processing of the contour surface extraction. The identification and reconstruction of experimental results from high-contrast and low-contrast objects agree well with the real objects. The exploration of the all-optical object identification and 3D reconstruction techniques provides potential applications of high efficiencies, low consumption, and compact systems.

Spiking neural networks (SNNs) utilize brain-like spatiotemporal spike encoding for simulating brain functions. Photonic SNN offers an ultrahigh speed and power efficiency platform for implementing high-performance neuromorphic computing. Here, we proposed a multi-synaptic photonic SNN, combining the modified remote supervised learning with delay-weight co-training to achieve pattern classification. The impact of multi-synaptic connections and the robustness of the network were investigated through numerical simulations. In addition, the collaborative computing of algorithm and hardware was demonstrated based on a fabricated integrated distributed feedback laser with a saturable absorber (DFB-SA), where 10 different noisy digital patterns were successfully classified. A functional photonic SNN that far exceeds the scale limit of hardware integration was achieved based on time-division multiplexing, demonstrating the capability of hardware-algorithm co-computation.

Extreme Learning Machines (ELMs) are a versatile Machine Learning (ML) algorithm that features as the main advantage the possibility of a seamless implementation with physical systems. Yet, despite the success of the physical implementations of ELMs, there is still a lack of fundamental understanding in regard to their optical implementations. In this context, this work makes use of an optical complex media and wavefront shaping techniques to implement a versatile optical ELM playground to gain a deeper insight into these machines. In particular, we present experimental evidences on the correlation between the effective dimensionality of the hidden space and its generalization capability, thus bringing the inner workings of optical ELMs under a new light and opening paths toward future technological implementations of similar principles.

Orbital angular momentum (OAM), emerging as an inherently high-dimensional property of photons, has boosted information capacity in optical communications. However, the potential of OAM in optical computing remains almost unexplored. Here, we present a highly efficient optical computing protocol for complex vector convolution with the superposition of high-dimensional OAM eigenmodes. We used two cascaded spatial light modulators to prepare suitable OAM superpositions to encode two complex vectors. Then, a deep-learning strategy is devised to decode the complex OAM spectrum, thus accomplishing the optical convolution task. In our experiment, we succeed in demonstrating 7-, 9-, and 11-dimensional complex vector convolutions, in which an average proximity better than 95% and a mean relative error <6 % are achieved. Our present scheme can be extended to incorporate other degrees of freedom for a more versatile optical computing in the high-dimensional Hilbert space.