Engineering nonlinear activation functions for all-optical neural networks via quantum interference.

Artificial neural networks (ANNs) have now been widely used for industry applications and also played more important roles in fundamental researches. Although most ANN hardware systems are electronically based, optical implementation is particularly attractive because of its intrinsic parallelism and low energy consumption. Here, we propose and demonstrate fully-functioned all optical neural networks (AONNs), in which linear operations are programmed by spatial light modulators and Fourier lenses, and optical nonlinear activation functions are realized with electromagnetically induced transparency in laser-cooled atoms. Moreover, all the errors from different optical neurons here are independent, thus the AONN could scale up to a larger system size with final error still maintaining in a similar level of a single neuron. We confirm its capability and feasibility in machine learning by successfully classifying the order and disorder phases of a typical statistic Ising model. The demonstrated AONN scheme can be used to construct various ANNs of different architectures with the intrinsic parallel computation at the speed of light.

The network coding paradigm enables better resource utilization and improves throughput in communication networks, thus, employing network coding in all-optical multicast networks will facilitate the future development of effective optical networks. However, some properties of optical networks, such as limited buffering and weak logic operation capabilities, restrict the application of network coding in optical multicast networks. So in this paper, several problems of adapting network coding to all-optical multicast networks are addressed: 1) Put forward a novel strategy to distribute encoding vectors in all-optical networks, and, 2) design a framework of Encoding Node Model to support the special scheme of encoding vectors assignment in practical networks. Moreover, the experimental demonstration indicates that it is feasible and efficient to employing network coding in optical-layer multicast traffic.

This article describes the results from the research work on design of a ultra low power off-line power supply with very high conversion efficiency. The input voltage is 230 VAC nominal and output voltage is 5 VDC. By ultra low power levels, an output power level in the area ranging from 50 mW and up to 1000 mW is meant. The small power supply is intended for use as a standby power supply in mains operated equipment, which requires a small amount of power in standby mode.

In this paper, we demonstrate excitability in a silicon-on-insulator microring. Neglecting the fast light dynamics simplifies the calculation of the excitability onset. Excitable microrings will be useful as a basic building block in all-optical neural networks.

The important research content of modern communication systems is to realize high-speed, stable, and intelligent information transmission and processing. All-optical neural networks based on the silicon integrated technology and phase change materials (PCMs) can realize picosecond-level modulation speed, faster processing speed, and lower energy consumption compared with the traditional electrical communication system. The photonic synapse is the core component of the all-optical neural network module, but the existing photonic synapses based on PCMs require a modulation distance (MD) of several micrometers or even ten micrometers to achieve a large output intensity range. In this paper, we propose an ultra-compact nonvolatile photonic synapse, in which MD can be shortened to 1 μm, breaking the record of the minimum signal MD of the silicon photonic synapse using the PCMs. At the same time, the output intensity range of our synapse is almost twice that of the existing research. Based on this photonic synapse, we analyze the relationship between the output response and incident wavelength, which can help to design an ultra-compact photonic convolutional neural network. This work has great potential in future photonic computing and photonic communication technologies.

Antimony trisulfide (Sb2S3) is an Earth abundant material that is transparent to visible and near infrared light (N-IR). Switching the material between amorphous and crystalline states causes radical property changes that deem it useful for programming the response of visible and N-IR photonics devices. We have demonstrated how Sb2S3 can be used to program high resolution micro-displays, dielectric metasurfaces, hyperbolic metamaterials, waveguides, and all-optical neural networks. This presentation will discuss these demonstrations and the underlying physics responsible for the phase transition in Sb2S3.

While the data throughput of electronic processors is rather high (TFLOP range), the question arises whether optical neural networks (NN) enable a value proposition for zero-delay (real-time) processing? The answer might be ‘yes’, since once the network is trained, the time to perform an inference tasks is simply given by the time-of-flight of the photon through the processor’s NN. Here we discuss how the three functions of the perceptron (dot-product synaptic weighting, summation, and nonlinear thresholding) can be mapped onto a) optoelectronic [George et al. Opt.Exp. 2019], and b) all-optical hardware [Miscuglio et al. OMEx 2018]. The latter is realized via co-integration of phase-change-materials atop Silicon photonics. Once trained, the weights only require rare updating, thus saving power. Performance wise, such an integrated all-optical NN is capable of < fJ/MAC using experimental demonstrated pump-probe [Waldecker et al, Nat. Mat. 2015] with a delay per perceptron being ~ps.

Single solid-state optical emitters have quantum mechanical properties that make them suitable for applications in information processing and sensing. Most of these quantum technologies rely on the capability to integrate the emitters in reliable solid-state optical networks. In this paper, we present integrated devices based on GaAs photonic crystals and InAs self-assembled quantum dots. These quantum networks are well suited to future optoelectronic devices operating at ultralow power levels, single-photon logic devices and quantum information processing.

We demonstrate parametric wavelength conversion via four-wave mixing using ultra-low peak pump powers of a few milliwatts in a micrometer-scale silicon device. The response time of our device is 100 ps allowing for implementation in high-bandwidth optical communications. At these ultra-low power levels and microscale sizes, it should be possible to realize hundreds of these devices operating simultaneously on a single chip.

The noise and gain for an InP HEMT were investigated as a function of DC power consumption. A special chip carrier in mixed coplanar waveguide/microstrip technology was developed for easy handling of the chip and to minimize the mounting parasitics. A complete S parameter and noise parameter model for the transistor and the carrier have been extracted. Measurements of the transistor mounted on the carrier gave an associated gain of 10 dB and noise figure lower than 0.7 dB at a DC power level of 200 /spl mu/W at 10 GHz. An X band one-stage amplifier with low DC consumption based on this transistor was built. The amplifier had a gain of 8-10 dB for the frequency range 8-12 GHz at V/sub ds/=0.35 V, I/sub ds/=3.9 mA. The noise figure of the complete amplifier was measured to 1.6 dB at 10 GHz. The HEMT was manufactured in our own laboratory process.

Young’s double-slit experiment played a crucial role in establishing the wave nature of light. In this chapter, the shocking result that incident electrons yield a similar interference pattern as that formed by light waves is described. It is shown that the only way the experimental results could be explained is via a wave function description of electrons. It is also shown that, in the same experiment, the interference fringes disappear if the which-path information becomes available. This is the essence of wave–particle duality. The first of the Einstein–Bohr debates on wave-particle duality and Bohr’s principle of complementarity in the double-slit experiment is also discussed. Also presented are the counterintuitive notions of delayed choice and quantum eraser effects showing how the availability or erasure of information generated in the future can affect how the data in the present can be interpreted.

Homomorphic encryption (HE) has emerged as a pivotal technology for secure neural network inference (SNNI), offering privacy-preserving computations on encrypted data. Despite active developments in this field, HE-based SNNI frameworks are impeded by three inherent limitations. Firstly, they cannot evaluate non-linear functions such as ReLU, the most widely adopted activation function in neural networks. Secondly, the permitted number of homomorphic operations on ciphertexts is bounded, consequently limiting the depth of neural networks that can be evaluated. Thirdly, the computational overhead associated with HE is prohibitively high, particularly for deep neural networks. In this paper, we introduce a novel paradigm designed to address the three limitations of HE-based SNNI. Our approach is an interactive approach that is solely based on HE, called iLHE. Utilizing the idea of iLHE, we present two protocols: ReLU, which facilitates the direct evaluation of the ReLU function on encrypted data, tackling the first limitation, and HeRefresh, which extends the feasible depth of neural network computations and mitigates the computational overhead, thereby addressing the second and third limitations. Based on HeReLU and HeRefresh protocols, we build a new framework for SNNI, named HeFUN. We prove that our protocols and the HeFUN framework are secure in the semi-honest security model. Empirical evaluations demonstrate that HeFUN surpasses current HE-based SNNI frameworks in multiple aspects, including security, accuracy, the number of communication rounds, and inference latency. Specifically, for a convolutional neural network with four layers on the MNIST dataset, HeFUN achieves 99.16% accuracy with an inference latency of 1.501 s, surpassing the popular HE-based framework CryptoNets proposed by Gilad-Bachrach, which achieves 98.52% accuracy with an inference latency of 3.479 s.

Deep learning embodied some essence of artificial intelligence, but it relied on data set and lacked migration learning ability. We should build general theorem to explain artificial intelligence from nature or human. We can treat each static data as a variable like wave-particle duality, then we can adopt idea from convolutional neural network or other machine learning algorithms to extract features from little data. This method can open up a new theory to accomplish migration learning and artificial intelligence. The theory will be supported by balance rule: everything has a gradient and tends to remain a zero gradient state, and we can connect different feature spaces by gradients.

Quantum mechanics (QM) has long challenged our understanding of time, space, and reality, with phenomena such as superposition, wave–particle duality, and quantum entanglement defying classical notions of causality and locality. Despite the predictive success of QM, its interpretations—such as the Copenhagen and many-worlds interpretations—remain contentious and incomplete. This paper introduces Strip Theory, a novel framework that reconceptualises time as a two-dimensional manifold comprising foretime, the sequential dimension, and sidetime, an orthogonal possibility dimension representing parallel quantum outcomes. By incorporating sidetime, the theory provides a unified explanation for quantum superposition, coherence, and interference, resolving ambiguities associated with wavefunction collapse. The methods involve extending the mathematical formalism of QM into a five-dimensional framework, where sidetime is explicitly encoded alongside spatial and sequential temporal dimensions. The principal findings demonstrate that this model reproduces all measurable results of QM while addressing foundational issues, offering a clearer and more deterministic interpretation of quantum phenomena. Furthermore, the framework provides insights into quantum coherence, wave–particle duality, and the philosophical implications of free will. These results suggest that Strip Theory can serve as a bridge between interpretations and provide a deeper understanding of time and reality, advancing both theoretical and conceptual horizons.

Space operation tasks can be significantly improved and optimised if they are provided with valuable supporting tools such as diagnostic systems, expert sys tems, smart monitoring and alarming tools or data-driven forecasting systems. This work describes three projects where soft computing techniques such as genetic algorithms, fuzzy logic and artificial neural networks have been used to provide the desired features. The obtained results are very promising and therefore the use of soft computing techniques is highly recommended for future projects. INTRODUCTION Some space operations tasks require advanced features that are difficult to provide using traditional methodologies. These advance methodologies include among others: • Find feasible solutions of a highly constrained problem. Example: unload bias manoeuvre in three-wheels stabilized spacecraft with many attitude changes. It is difficult to find a solution that does not violate the saturation, stiction, stability or feasibility constraints. • Optimisation techniques can help to find a good solution in fixed time. Example: optimise the fuel spent in the unload bias manoeuvre taking into account the next orbit plans and constraints before the next orbit starts. • Monitoring: a method that combines several telemetry parameters and provides gradual alarms is very convenient as decision support tool. Example: monitor the spacecraft’s gyroscope fault detection; once a problem is detected a gradual alarm is fired. The engineers have time to react. • Diagnosis: the diagnosis goes beyond the monitoring. It consists of knowing the cause of an unexpected behaviour. Example: explain why an alarm has been triggered. • Data-Driven Modelling: it only takes into account the real behaviour of the spacecraft in space. By this way assumptions that are usually done when building physical models are avoided. The resulting model can be used to apply “what-if” analysis. Example: build a virtual thermal sensor to recover a failed one using only spacecraft data. SOFT COMPUTING TECHNIQUES Soft computing techniques encompass: • Genetic Algorithms: robust search algorithms that do not require knowledge of the objective function to be optimised and can search through large spaces quickly. They were derived from processes of molecular biology and the evolution of life. Their operators – crossover, mutation, and reproduction – match the synonymous biological processes. Instead of DNA, they usually process strings of symbols of finite length; these symbols encode the parameters to be optimised. • Fuzzy Logic: it provides an approach to approximate reasoning in which the rules of inference are approximate rather than exact. Fuzzy logic is useful in manipulating information that is incomplete, imprecise, or unreliable. Also called fuzzy set theory, fuzzy logic extends the simple Boolean operators, can express implication, and is used extensively in Artificial Intelligence (AI) programs. • Artificial Neural Networks: are parallel computation models made of adaptative processing units called neurons. These networks are parallel implementations of non-linear systems. Their major characteristic is the ability to learn by examples. This ability is very useful when we do not know how a complex system works but we have some data (examples) of its behaviour. Another interesting characteristic is its intrinsic parallelism that allows obtaining solutions very fast when implemented in hardware.