Abstract The energy demanding nature of deep learning (DL) has fueled the immense attention for neuromorphic architectures due to their ability to operate in a very high frequencies in a very low energy consumption. To this end, neuromorphic photonics are among the most promising research directions, since they are able to achieve femtojoule per MAC efficiency. Although electrooptical substances provide a fast and efficient platform for DL, they also introduce various noise sources that impact the effective bit resolution, introducing new challenges to DL quantization. In this work, we propose a quantization-aware training method that gradually performs bit reduction to layers in a mixed-precision manner, enabling us to operate lower-precision networks during deployment and further increase the computational rate of the developed accelerators while keeping the energy consumption low. Exploiting the observation that intermediate layers have lower-precision requirements, we propose to gradually reduce layers’ bit resolutions, by normally distributing the reduction probability of each layer. We experimentally demonstrate the advantages of mixed-precision quantization in both performance and inference time. Furthermore, we experimentally evaluate the proposed method in different tasks, architectures, and photonic configurations, highlighting its immense capabilities to reduce the average bit resolution of DL models while significantly outperforming the evaluated baselines.

We demonstrate experimentally a bandwidth-reconfigurable mmWave Fiber Wireless (FiWi) fronthaul bus topology for spectrally efficient and flexibly reconfigurable 5G Centralized-Radio Access Networks (C-RAN). The proposed fronthaul architecture includes four 1 Gb/s Intermediate Frequency over Fiber (IFoF) channels that can be flexibly allocated among two in-series Reconfigurable Optical Add/Drop Multiplexer (ROADM) integrated nodes, supporting in total 8 V-band 32-element Phased Array Antenna (PAA) terminals. The ROADM was fabricated as an integrated photonic device exploiting the ultra-low loss Si3N4TriPleX waveguide integration platform and an architectural layout based on cascaded MZI interleavers. The device has flat top response of 32.5 GHz with a Free Spectral Range (FSR) of 100 GHz and fiber-to-fiber losses of 5 dB, while the V-band PAA supports analog RF beamsteering capabilities within a 90°-sector and 1m wireless distance. Each of the FiWi links carries a 250 MBd QAM16 waveform enabling a total of 1 Gb/s rate per end user beam, complying with the 5G Key Performance Indicator (KPI) user-rate requirement. Bandwidth-reconfigurability is experimentally demonstrated by selectively dropping channels either at the first or at the second ROADM node, allowing in this way the bandwidth allocation to be flexibly defined between two different network segments. Both uplink and downlink performance are experimentally validated for different ratios of bandwidth allocation among the two nodes, revealing Error Vector Magnitude (EVM) values that meet the respective 3GPP signal quality specifications. The two-stage FiWi IFoF/mmWave fronthaul bus topology, based on a miniaturized, integrated, low loss Si3N4ROADM and supporting high-capacity wireless beamsteering capability can form a promising roadmap towards flexible and reconfigurable 5G C-RAN architectures.

Photonic neuromorphic hardware can provide significant performance benefits for Deep Learning (DL) applications by accelerating and reducing the energy requirements of DL models. However, photonic neuromorphic architectures employ different activation elements than those traditionally used in DL, slowing down the convergence of the training process for such architectures. An initialization scheme that can be used to efficiently train deep photonic networks that employ quadratic sinusoidal activation functions is proposed in this paper. The proposed initialization scheme can overcome these limitations, leading to faster and more stable training of deep photonic neural networks. The ability of the proposed method to improve the convergence of the training process is experimentally demonstrated using two different DL architectures and two datasets.