Neuroai Advances AI Efficiency Via Neuroscience, Findings from August 2025 Workshop

Scientists are increasingly recognising the untapped potential arising from a closer integration of neuroscience and artificial intelligence (AI). Jean-Marc Fellous, Gert Cauwenberghs, and Cornelia Fermüller, from the University of California, San Diego and University of Maryland respectively, alongside Yulia Sandamisrkaya and Terrence Sejnowski et al., outline a vision for ‘NeuroAI’ , an AI approach informed by the principles of biological neural systems. This work, stemming from a workshop held in August 2025, identifies key areas where these fields can synergise, including embodiment, language and robotics, and argues that NeuroAI could not only enhance the efficiency of AI algorithms but also fundamentally reshape our understanding of the brain itself.

These programs would encourage postdoctoral fellows and graduate students to spend two to three months in different laboratories, experiencing diverse theoretical perspectives and methodologies to strengthen trans-disciplinary research. Researchers recognised the scale of these efforts often exceeds single laboratory capacity, suggesting larger collaborative initiatives mirroring approaches used in particle physics and astronomy, such as those employed at CERN and with the Webb telescope. The work championed a broad range of approaches, extending beyond bio-inspired designs to include utilising biological materials and abstract biological concepts for novel device creation. Furthermore, the study stressed the importance of establishing ethical and legal standards alongside the advancement of embodied cognition and computation. The team posited that embodiment is a necessary, though not sufficient, condition for creating generally intelligent agents, drawing on the observation that all known intelligent agents are embodied. Experiments explored the limitations of current Large language models (LLMs) and Large reasoning models (LRMs), noting their cognitive fragility and lack of biological plausibility despite achieving human or super-human performance on specific language tasks. Scientists proposed improving these models by incorporating insights from the efficiency of biological brains to advance both scientific understanding of human minds and improve human lives. The study questioned how to build more cognitively plausible AI models, specifically addressing the degree of modularity needed between language and reasoning systems and the role of different learning systems, such as the hippocampal and basal ganglia systems, in chain-of-thought reasoning and consciousness. Researchers also investigated how to create developmentally plausible models, drawing parallels to infant language acquisition and the potential for incorporating innate mental structures as scaffolding for AI development.,

Smartphone NPUs enable wider NeuroAI applications

Researchers recorded that worldwide smartphone sales exceed $1.2 billion per year, with 9 billion smartphones currently in use, positioning smartphone Neural Processing Units (NPUs) as leading technology in battery-powered AI processing. The team measured that these NPUs offer peak cost, power, and throughput, making them suitable for applications beyond smartphones, including robotics, IoT, and prosthetics. Experiments revealed a critical need for accessible DRAM interface technology to scale AI model development, as co-design of hardware and algorithms has been limited by memory capacity. Current neuromorphic accelerator chips, while massively parallel, are restricted to small-scale recognition tasks due to reliance on SRAM cells, which cost over 20times more per bit than DRAM.

Data shows that as state-of-the-art models scale to billions of parameters, translation to SRAM memories is impractical due to area constraints, necessitating commensurate memory capacity achievable with DRAM technology at approximately 1/100 the cost. Concerns were raised regarding the complexity and expense of DRAM interface IPs, highlighting the need for joint open-source efforts. Measurements confirm that DRAM memory achieves high capacity through 3D monolithic technologies stacking over 200 layers. The study advocates for tightly integrating compute, communication, and memory within this 3D framework, coupled with algorithms enforcing bit-level, modular, and temporal sparsity, to minimize area and energy footprints.

Researchers project that within 0.5 years, a neuromorphic twin derived from a fully-mapped mouse connectome will be achievable, alongside the incorporation of biological neurons, synapses, and modulatory computation into AI systems. Furthermore, integration of 3D memories supporting sparse workloads, open access to memory and sensors, and DRAM chip interface IP blocks for academic scaling are anticipated within this timeframe. The work details that FeFET, RRAM, PCM, and MRAM technologies will enable novel Compute-In-Memory (CIM) and Processing-In-Memory (PIM) architectures. Scientists envision neuromorphic deep brain stimulation technology and wearable devices incorporating neuromorphic computing, such as glasses, earables, and health monitors, becoming reality. The research also highlights the potential for foveated projections, generative event-based/low-power AR/XR, and AI-driven programming of micro-code generation for neuromorphic computing, alongside the opening of powerful smartphone NPUs to broader applications and the development of hardware RNNs and SSMs as community chip developments.

NeuroAI’s potential benefits and future challenges are significant

The resulting work identifies key areas where these fields can mutually benefit, including embodiment, language, robotics, and learning mechanisms. The workshop participants, including leading researchers, considered the potential of NeuroAI through SWOT analyses, highlighting both its opportunities and potential risks. Discussions extended beyond technical advancements to address broader implications for education and ethics, particularly concerning the future role of human learning if AI systems surpass human capabilities in various intellectual domains. Participants also contemplated the ethical considerations of interacting with AI systems exhibiting empathy and emotional responses, especially as embodiment becomes a central goal.

Acknowledging the rapid pace of progress, the authors suggest that achieving human-level intelligence and emotion in AI is no longer a distant fantasy. They note a limitation in addressing the ethical and educational challenges this presents, as current approaches often assume AI will remain a controllable tool. Future research should focus on proactively addressing these concerns and exploring the long-term societal impacts of increasingly sophisticated AI systems.