Neuromorphic-based metaheuristics: A new generation of low power, low latency and small footprint optimization algorithms

Prof. El-Ghazali Talbi · 2025

Read Neuromorphic-based metaheuristics: A new generation of low power, low latency and small footprint optimization algorithms by Prof. El-Ghazali Talbi (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-16362.txt.

Introduction

El-Ghazali Talbi argues that brain-like chips could run search algorithms on a sip of power. Metaheuristics are the workhorse tricks for hard optimization, from routing to scheduling. Neuromorphic hardware computes with spiking neurons instead of the usual von Neumann design, promising low energy and fast response.

Porting search methods to spikes is not straightforward, the paper stresses. It sorts existing attempts by algorithm family and problem type, and weighs what each gains or loses. The survey closes with open challenges for this young field, which it dubs Nheuristics. Cheap, tiny optimizers could end up inside sensors and robots everywhere.

Catalog record

CREATORProf. El-Ghazali Talbi
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE101 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-16362

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Neuromorphic-based metaheuristics: A new generation of low power, low latency and small footprint optimization algorithms Prof. El-Ghazali Talbi Abstract Neuromorphic computing (NC) introduces a novel algorithmic paradigm representing a major shift from traditional digital computing of Von Neumann architectures. NC emulates or simulates the neural dynamics of brains in the form of Spiking Neural Networks (SNNs). Much of the research in NC has concentrated on machine learning applications and neuroscience simulations. This paper investigates the modelling and implementation of optimization algorithms and particularly metaheuristics using the NC paradigm as an alternative to Von Neumann architectures, leading to breakthroughs in solving optimization problems. Neuromorphic-based metaheuristics (Nheuristics) are supposed to be characterized by low power, low latency and small footprint. Since NC systems are fundamentally different from conventional Von Neumann computers, several challenges are posed to the design and implementation of Nheuristics. A guideline based on a classification and critical analysis is conducted on the different families of metaheuristics and optimization problems they address. We also discuss future directions that need to be addressed to expand both the development and application of Nheuristics. This work has been supported by the ERC Generator of the University of Lille. 1 Introduction On one hand, digital computing based on clock-driven architectures faces major barriers due to the end of Dennard scaling [ 1 ] , the slowing of Moore’s Law [ 2 ] , and the limited bandwidth between CPU and memory—known as the Von Neumann bottleneck or memory wall. These issues result in higher energy consumption, increased latency, and larger hardware footprints. On the other hand, the demand for solving large-scale optimization problems—such as training deep neural networks, large language models, and complex engineering systems—continues to grow, requiring faster and more energy-efficient solutions. As a result, there is an urgent need for novel computing paradigms and optimization methods that minimize power usage, latency, and size. The rapid, often unsustainable growth of AI is also driving up carbon emissions; for instance, training a single large model can emit over 284 tons of CO2 [ 3 ] [ 4 ] , and AI could account for 0.5% of global electricity use by 2027 [ 5 ] . Moreover, applications such as edge computing and IoT (Internet of Things), where low latency and small form factors are critical, further underscore the need for more power efficient alternatives [ 6 ] . Neuromorphic Computing (NC) is an emerging and promising alternative to traditional Von Neumann architectures [ 7 ] [ 8 ] . Inspired by the structure and function of the human brain, NC aims to develop new hardware and algorithms that enable next-generation, non-digital computing paradigms. These systems are designed to provide significant improvements in energy efficiency, reduced latency, and smaller physical footprints (including size and weight). The NC field is inherently interdisciplinary, bringing together researchers from materials science, neuroscience, electrical engineering, and computer science. NC has been proven to be Turing-complete, making it capable of performing general-purpose computing [ 9 ] [ 10 ] . Most of the research work in NC has focused on the development of cognitive applications (e.g., machine learning, computational neuroscience). The application of NC extends well beyond cognitive tasks such as graph algorithms, random walks, partial differential equation (PDE) solving and signal processing [ 11 ] . These applications highlight the versatility and potential of NC, offering new avenues for solving complex computational problems across various disciplines. In this paper, we explore the use of NC for the design and implementation of optimization algorithms, with a particular focus on neuromorphic-based metaheuristics (Nheuristics) for tackling challenging optimization problems. NC opens up a new algorithmic landscape for optimization, offering a multi-layered approach that integrates hardware and algorithms to achieve brain-like efficiency. We believe there is a significant algorithmic opportunity for Nheuristics. However, since NC systems are fundamentally different from traditional Von Neumann computers, designing and implementing Nheuristics presents several challenges. We propose a unified framework for defining the key concepts necessary for the design and implementation of Nheuristics. Given the characteristics of Spiking Neural Networks (SNNs) and neuromorphic hardware, new algorithmic directions are required to leverage their implicit recurrence, event-driven nature, and sparse computational features. Novel research is needed to develop neuron models that can process information over time, devise spike-based encodings to represent solutions and data, and design SNN architectures and learning rules suited for sparse, asynchronous, and event-driven dynamical systems. The paper is structured as follows. In section 2 (resp. 3 ), the main concepts of NC (resp. metaheuristics and optimization problems) are provided. Section 4 (resp. 5 ) defines a general and unified framework for the design (resp. implementation) of Nheuristics. Section 6 details a classification and a critical analysis of the different families of Nheuristics (i.e., greedy, local search based, evolutionary and swarm intelligence algorithms). Finally, the last section 7 presents the main conclusions and opens some important research perspectives for the future development of Nheuristics. 2 Neuromorphic computing In NC, the computational paradigm is inspired from the most efficient information processing system known to humanity: the brain [ 12 ] . NC, which emerged in the late 1980s, focuses on developing sensing and processing systems that mimic the dynamics of the brain [ 7 ] [ 8 ] . SNNs represent the main algorithmic paradigm for NC that attempts to mimic the synaptic functionalities of the neurons in a temporal, distributed and asynchronous event-driven way. They are considered as the emerging third generation of the neural network model [ 13 ] [ 14 ] [ 15 ] . Unlike Von Neumann architectures where programs are defined by explicit instructions, programs in NC are defined by the structure of the SNN and its parameters. SNNs use spikes to encode, process and communicate information. The characteristics of those spikes (e.g., time at which they occur, magnitude, shape) will allow to encode numerical information, while in Von Neumann architectures, the information is encoded as numerical values. Another important property of SNNs is their connection plasticity. The synaptic weight between two neurons can be adjusted which enables the SNNs to model and solve complex computational problems. NC presents some fundamental operational differences with digital computing: 1. Energy efficiency : The brain manages billions of neurons and trillions of synapses requiring about 20 watts of power [ 16 ] , while training on a digital computer a large language model (LLM) like GPT-3 requires 1,300 megawatt hours (MWh) of electricity. NC systems are designed to operate on orders of magnitude less power consumption than traditional CPUs or GPUs, making them suitable for battery-operated devices and large-scale implementations. Neuromorphic chips can achieve performance with power consumption as low as milliwatts, compared to watts for conventional systems. Just a single modern GPU such as an NVIDIA Tesla V100, requires about 300 watts. The recent Exascale digital machines are expected to draw tens to hundreds of Megawatts of power. 2. Massively inherent parallelism : Neuromorphic computers capitalize on their inherent parallelism, with a large number of neurons and synapses capable of operating in a parallel way. The number of neurons in the human brain is in the order of 86 billion. Each neuron can establish thousands of synaptic connections with other neurons, amounting to approximately 10^{15}

Read the complete text