Biologically plausible energy-efficient context-sensitive neural networks

The rapid advancements in artificial intelligence (AI), particularly in deep neural networks (DNNs), have significantly driven the demand for specialised hardware, such as GPUs and TPUs, to meet the growing computational requirements. However, this surge in computational power has led to substantial increases in energy consumption, raising critical concerns about the economic, technical, and environmental sustainability of AI technologies. For instance, training a single large-scale neural network can generate as much carbon dioxide as five automobiles over their lifespans. While notable efforts have been made to enhance the energy efficiency of deep learning (DL) models, the challenge of sustainable computation persists, especially in resource-constrained environments like hearing aid devices. Despite being inspired by the human brain’s functionality, DNNs fall short of matching the brain’s remarkable energy efficiency, where the human brain operates at merely 20 watts while performing complex cognitive tasks. This thesis addresses the imperative need for human-level computational efficiency in DL models through a multidisciplinary approach synthesising AI, neuroscience, and hardware engineering insights. Central to this work is integrating concepts derived from recent neurological discoveries of context-sensitive neurons into DNNs, aiming to enhance energy efficiency in neural network processing. The research introduces a novel context-sensitive (CS) mechanism within a deep convolution network (DNN), demonstrating significant reductions in neural activity and energy consumption compared to state-of-the-art methods. The CS DNN with 18 convolution layers is employed for multimodal (MM) non-linear regression tasks, such as audio-visual speech enhancement (AVSE). It is then compared against a conventional point neuron-inspired DCN in terms of perceptual evaluation of speech (PESQ) and short-time objective intelligibility (STOI). This research shows that the two-point neuron-driven DCN performs comparable to point-neuron DCN. However, the twopoint neuron uses up to 7 times fewer neurons. Furthermore, to demonstrate the scalability of this energy efficiency, the DCN is mapped to a 50-layer convolution network and implemented on Xilinx UltraScale+ MPSoC based Open-MHA (Multimodal hearing aid) platform. The two-point neuron architecture showed 103 times lesser energy (J) consumption for an inference. This research introduces a novel context-sensitive (CS) mechanism within a deep convolutional network (DNN), demonstrating significant reductions in neural activity and energy consumption compared to state-of-the-art methods. The proposed CSDNN, comprising 18 convolutional layers, is employed for multimodal non-linear regression tasks such as audio-visual speech enhancement (AVSE). Its performance is evaluated against a conventional point-neuron-inspired deep convolutional network (DCN) using the Perceptual Evaluation of Speech Quality (PESQ) and Short-Time Objective Intelligibility (STOI) metrics. The results indicate that the two-point neuron-driven DCN performs comparably to the point-neuron DCN while utilizing up to seven times fewer neurons. To demonstrate the scalability of this energy-efficient architecture, the DCN is expanded to a 50-layer convolutional network and implemented on the Xilinx UltraScale+ MPSoC-based Open-MHA (Open Multimodal Hearing Aid) platform. The two-point neuron architecture exhibited a 103-fold reduction in energy consumption (J) per inference. Furthermore, this research adopts a two-point neuron-based, biologically plausible training mechanism, which is transformed into a novel multimodal setting and applied to the audio-visual speech enhancement task. Experimental results, compared to a backpropagation-based baseline model, demonstrate outstanding energy efficiency, reducing neuron firing rates by up to 70%. This reduction implies more sustainable implementations, making the approach highly suitable and desirable for embedded systems. Finally, this thesis explores the biological plausibility of the proposed mechanism by implementing context-sensitive (CS) spiking neurons within a spiking neural network (SNN). This implementation provides a comprehensive understanding of the role of two-point neurons both mathematically and empirically. The CS-SNN is applied to a classical non-linear XOR learning task, demonstrating rapid learning—twice as fast as the baseline model—along with improved performance. Building on these contributions, this research contributes to the ongoing pursuit of sustainable and biologically inspired AI by proposing and validating a context-sensitive approach that advances the energy efficiency of deep neural networks (DNNs).

Citation

Naseem, M.R. (2025) Biologically plausible energy-efficient context-sensitive neural networks. University of Wolverhampton. https://wlv.openrepository.com/handle/2436/625937

Publisher

University of Wolverhampton

URI

https://wlv.openrepository.com/handle/2436/625937

Type

Thesis or dissertation

Language

en

Description

A thesis submitted in partial fulfilment of the requirements of the University of Wolverhampton for the degree of Doctor of Philosophy.

Collections

Theses and Dissertations