Integrated Optimization and Hardware-Algorithm Co-Design in Neural Networks

The recent advancements in neural network optimization and hardware-algorithm co-design have significantly pushed the boundaries of efficiency and performance in resource-constrained environments. A notable trend is the integration of deep reinforcement learning for optimizing core placement in many-core near-memory computing systems, which aims to enhance parallelism and reduce power consumption. Additionally, the development of binary-native and gradient-free training algorithms for binary neural networks has opened new avenues for operation-optimized training, demonstrating substantial accuracy improvements with minimal hardware requirements. Hardware-aware training methodologies are also gaining traction, particularly in optimizing memory usage for neural networks deployed on event-based processors, where routing-aware training has shown to significantly enhance both accuracy and memory efficiency. Furthermore, the field is witnessing innovative approaches to error detection and correction in ReRAM crossbar arrays, ensuring fault-free accuracy in deep learning accelerators with minimal overhead. These developments collectively underscore a shift towards more integrated and efficient solutions that bridge the gap between algorithm design and hardware implementation, promising scalable and robust neural network deployments in diverse computational environments.

Sources

Core Placement Optimization of Many-core Brain-Inspired Near-Storage Systems for Spiking Neural Network Training

Training Multi-Layer Binary Neural Networks With Local Binary Error Signals

Hardware architecture and routing-aware training for optimal memory usage: a case study

Estimation during Design Phases of Suitable SRAM Cells for PUF Applications Using Separatrix and Mismatch Metrics

Online Soft Error Tolerance in ReRAM Crossbars for Deep Learning Accelerators

BinSparX: Sparsified Binary Neural Networks for Reduced Hardware Non-Idealities in Xbar Arrays

Built with on top of