19 August 2025 in IEEE Access
This paper presents BAM-Net, a hardware-efficient binarization algorithm designed for associative memory (AM) implementation. BAM-Net aims to reduce memory overhead, power consumption, and implementation complexity. Motivated by the observation that the largest positive weight in each row of a Hebbian-trained weight matrix dominates the recall process, the proposed approach retains only this maximum weight per row. This yields a highly sparse binary matrix containing only 0 and +1 values. The resulting simplification enables the exclusive use of standard NAND gates in hardware implementation, thereby eliminating the need for multi-level weight storage or complex arithmetic operations. Experimental results on synthetic 7×5 pattern recall and the binarized MNIST dataset demonstrate that BAM-Net achieves superior recall accuracy, up to 96.2% and 92.7% respectively, while reducing memory usage by over 50% compared to state-of-the-art quantum, photonic, and quantized AM models. The proposed method converges in fewer recall attempts and exhibits excellent scalability, making it well suited for low-power, real-time, and resource-constrained hardware applications.
1 December 2025 in Photonics Research
The exponential growth in data traffic has driven significant research into maximizing the capacity of free-space optical (FSO) communication systems. Orbital angular momentum (OAM) multiplexing offers a promising approach by using spatially structured beams with helical wavefronts to achieve higher data transmission rates. However, conventional electronic convolutional-neural-network-based OAM demultiplexing schemes exhibit substantial computational and energy efficiency limitations. In this paper, we introduce a hybrid optical-electronic Fourier phase shift neural network that implements phase-only feature extraction of the input multiplexed OAM beams in the Fourier domain. The proposed hybrid neural network uses phase spatial frequency kernels with the spatial light modulator to perform additive phase modulation of the Fourier-transformed input beams. Experimental results show that the proposed phase shift neural network has 6.5 times faster training time and three orders of magnitude higher energy efficiency compared to the designed conventional all-electronic convolutional neural network with one single convolution layer. The proposed system represents an idea towards energy-efficient, high-throughput optical neural networks for OAM-based FSO communication systems.