Freely scalable and reconfigurable optical hardware for deep learning

As deep neural network (DNN) models grow ever-larger, they can achieve higher accuracy and solve more complex problems. This trend has been enabled by an increase in available compute power; however, efforts to continue to scale electronic processors are impeded by the costs of communication, thermal management, power delivery and clocking. To improve scalability, we propose a digital optical neural network (DONN) with intralayer optical interconnects and reconfigurable input values. The path-length-independence of optical energy consumption enables information locality between a transmitter and a large number of arbitrarily arranged receivers, which allows greater flexibility in architecture design to circumvent scaling limitations. In a proof-of-concept experiment, we demonstrate optical multicast in the classification of 500 MNIST images with a 3-layer, fully-connected network. We also analyze the energy consumption of the DONN and find that digital optical data transfer is beneficial over electronics when the spacing of computational units is on the order of $$>10\,\upmu $$ > 10 μ m.

Minimization Number of Network-Coded Links Based on Improved Adaptive Genetic Algorithm for Multi-source Optical Networks

With the rapid development of multi-source optical multicast application, the wavelength division multiplexing (WDM) with limited number of wavelength channels is facing with the new challenge of bandwidth shortage. Optical multicast adopting network coding can improve the bandwidth utilization, but optical network coding needs to increase optical storage and computation overhead in WDM optical network. For reducing the number of optical network-coded links, an improved adaptive genetic algorithm (IAGA) is proposed to minimize the number of network-coded links for multicast. By designing the maximization difference crossover operation, IAGA can guarantee the diversity of population and avoid individuals from falling into a local optimal. By adaptively adjusting the crossover probability, IAGA makes the population diverse at the beginning stages and makes the excellent individuals remain in a stable condition. Compared with other algorithms, the simulation result shows that the proposed algorithm has fastest convergence speed, which means that it takes the shortest time to find the minimum numbers of coded link solutions.

Maximization Network Throughput Based on Improved Genetic Algorithm and Network Coding for Optical Multicast Networks

Maximal multicast stream algorithm based on network coding (NC) can improve the network’s throughput for wavelength-division multiplexing (WDM) networks, which however is far less than the network’s maximal throughput in terms of theory. And the existing multicast stream algorithms do not give the information distribution pattern and routing in the meantime. In the paper, an improved genetic algorithm is brought forward to maximize the optical multicast throughput by NC and to determine the multicast stream distribution by hybrid chromosomes construction for multicast with single source and multiple destinations. The proposed hybrid chromosomes are constructed by the binary chromosomes and integer chromosomes, while the binary chromosomes represent optical multicast routing and the integer chromosomes indicate the multicast stream distribution. A fitness function is designed to guarantee that each destination can receive the maximum number of decoding multicast streams. The simulation results showed that the proposed method is far superior over the typical maximal multicast stream algorithms based on NC in terms of network throughput in WDM networks.

A Grooming Nodes Optimal Allocation Method for Multicast in WDM Networks

The grooming node has the capability of grooming multicast traffic with the small granularity into established light at high cost of complexity and node architecture. In the paper, a grooming nodes optimal allocation (GNOA) method is proposed to optimize the allocation of the grooming nodes constraint by the blocking probability for multicast traffic in sparse WDM networks. In the proposed GNOA method, the location of each grooming node is determined by the SCLD strategy. The improved smallest cost largest degree (SCLD) strategy is designed to select the nongrooming nodes in the proposed GNOA method. The simulation results show that the proposed GNOA method can reduce the required number of grooming nodes and decrease the cost of constructing a network to guarantee a certain request blocking probability when the wavelengths per fiber and transmitter/receiver ports per node are sufficient for the optical multicast in WDM networks.

A Self-Driven and Adaptive Adjusting Teaching Learning Method for Optimizing Optical Multicast Network Throughput

With the development of one point to multiple point applications, network resources become scarcer and wavelength channels become more crowded in optical networks. To improve the bandwidth utilization, the multicast routing algorithm based on network coding can greatly increase the resource utilization, but it is most difficult to maximize the network throughput owing to ignoring the differences between the multicast receiving nodes. For making full use of the destination nodes’ receives ability to maximize optical multicast’s network throughput, a new optical multicast routing algorithm based on teaching-learning-based optimization (MR-iTLBO) is proposed in the paper. In order to increase the diversity of learning, a self-driven learning method is adopted in MR-iTLBO algorithm, and the mutation operator of genetic algorithm is introduced to prevent the algorithm into a local optimum. For increasing learner’s learning efficiency, an adaptive learning factor is designed to adjust the learning process. Moreover, the reconfiguration scheme based on probability vector is devised to expand its global search capability in MR-iTLBO algorithm. The simulation results show that performance in terms of network throughput and convergence rate has been improved significantly with respect to the TLBO and the variant TLBO.