Editorial: Brain-inspired computing: from neuroscience to neuromorphic electronics for new forms of artificial intelligence

摘要

The increasing diffusion of AI applications into daily life led to a significant rise in demand for advanced machine learning systems, such as artificial neural networks, which now outperform humans in many tasks. The rapid growth of generative AI solutions based on transformer architecture [1] has further accelerated the need for more powerful computational hardware. Additionally, research in humanoid robotics has focused on developing systems that replicate neural processes. However, conventional hardware solutions are unsustainable, as they require frequent training cycles, supervised learning, and large offline datasets which constrain the adoption of sustainable AI. Recently, industrial applications using conventional models have emerged, but neuromorphic approaches, inspired by the brain's functioning, offer promising, sustainable alternatives [2]. Neuromorphic is an umbrella term that spans many interdisciplinary fields, including neuroscience, material science and electronic architectures, extending into mathematical and software models. Advances in computational neuroscience, along with the development of neuronal and synaptic models have driven the emergence of neuro-inspired microelectronics. Firstly, the proposed circuits were primarily based on the observation that transistors operating in the sub-threshold regime share remarkable similarities with the biophysics of biological neuronal membranes [3]. This paved the way for the development of novel architectures based on silicon neurons. The maturity of the CMOS process allowed the steady implementation of brain-machine interfaces and neuro-inspired low-power computation systems, achieving higher levels of complexity [3]. However, more recently the scientific community acknowledged the superior performance of new materials and emerging devices in mimicking neuronal behaviors, further accelerating the research in this direction. Notable examples are functionalized nanomaterials [4][5] and memristive devices [6][7], which have demonstrated the ability to replicate synaptic plasticity through longand/or short-term changes in synaptic efficacy [8]. As these new solutions stabilize and move toward commercial viability, architectures based on them are emerging mainly in the form of Spiking Neural Networks (SNN) that outperform traditional platforms in distributed computation, showing higher energy efficiency [7]. Interestingly, an emerging domain of theoretical and computational neuroscience, based on a Bayesian approach adopted to model brain functions [9], recently opened promising perspectives in terms of energy efficient neuromorphic applications.Much of the focus in the development of neuromorphic solutions has been on the hardware. Conversely, on the software side, efforts were mainly aimed at creating AI algorithms inspired by neuronal architectures. Despite the recent increase in publications on AI solutions based on artificial neural network (ANN) and the recognized success of generative AI machineries, there is a growing consensus that alternative approaches must be investigated and implemented. This is due to the unsustainability of the current approach, that is evidently too resource-hungry (i.e., it is associated with unbearable energy and water consumption, as well as land use) [10]. In this respect, the brain, due to its event-based communication, remains the key model to emulate by virtue of its remarkable computational power despite its limited energy resources.In this fast-paced growth context, significant research efforts are often carried out within individual scientific domains. However, future breakthroughs are likely to come from cross-domain research encompassing many sectors such as neuroscience, electronics, computer science, and robotics, all driven by the same underlying goals and foundational principles. This joint Frontiers in Neuroscience and Frontiers in Electronics topic aims at showcasing the latest advancements in neuromorphic computing