FPGA-Based Training of Bio-Inspired Dendritic Neural Networks with Hebbian-Augmented Learning
Year: 2025
Type: preprint
Abstract: This work presents a bio-inspired artificial neural network (ANN) architecture that integrates dendritic structure and spatially constrained input sampling, mirroring key biological properties of neurons. The model combines traditional backpropagation with a novel, local Hebbian-like learning rule to improve both learning efficiency and biological plausibility. To address the energy demands of ANN training, the architecture is implemented on a Field Programmable Gate Array (FPGA), specifically the Zynq UltraScale+ ZCU102 board. FPGAs allow direct hardware mapping of algorithms, supporting highly parallel, low-latency computation with significant energy savings compared to conventional CPUs and GPUs. The proposed FPGA-based system significantly outperforms traditional platforms in both speed and efficiency. On the MNIST dataset, the FPGA executes one training epoch in just 13.46 seconds, compared to 480 seconds on a CPU and 45 seconds on a GPU. In terms of energy efficiency, the FPGA implementation achieves a 346-fold improvement over the CPU and a 57.5-fold improvement over the GPU. These results underscore the potential of combining bio-inspired ANN architectures with FPGA hardware to enable fast, energy-efficient neural computation, making them well-suited for embedded and edge AI applications.
Subfield: Hardware and Architecture
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W4415743577
Cites:
Cited by:
Related to: