Neuromorphic Architectures for Energy-Efficient Brain–Computer Interfaces

Modern AI algorithms do the opposite. Training a single large transformer can emit hundreds of thousands of pounds of CO₂ once tuning and experimentation are included (Strubell et al., 2019), and GPT-3's training run alone consumed over a million kilowatt-hours. The demand for frontier models grew roughly 300,000-fold between 2012 and 2018. Part of this is architectural: conventional von Neumann computing separates memory from processing, forcing constant data movement, a "memory wall" that dominates energy cost regardless of algorithmic tuning. Underneath that sits a physical floor: the Landauer limit on irreversible computation (the Landauer limit). 

Today's chips burn orders of magnitude more energy per operation than physics ultimately requires. Closing this gap is one of the defining problems at the intersection of neuroscience, semiconductor engineering, and machine learning. Non-invasive brain–computer interfaces (BCIs) offer a way to study this gap directly. Functional ultrasound (fUS) has matured quickly: closed-loop ultrasonic BMIs have decoded motor intent from primate cortex with stable performance across sessions spanning months (Griggs et al., Nature Neuroscience, 2024), and fUS has been demonstrated through intact human skulls (Norman et al., Science Translational Medicine, 2024). Yet current systems still rely on a conventional sensing–computation–actuation pipeline with continuous sampling, external processing, large-model decoding, stimulation. It introduces introduces latency, heat, and power draw incompatible with continuous wearable use.

The research explores hardware–software co-design to redesign this stack from the ground up, rather than optimizing sensing, compute, and learning separately. Two approaches tackle the memory wall head-on. The first is memristive crossbar arrays, chips that store and process information in the same spot instead of shuttling it back and forth between separate parts. That means synapse-like operations use much less energy than they would on standard digital chips (Choi et al., Nature Communications, 2022). The second: spiking neural networks, which only activate when there's something worth responding to, the same phenomenon the brain uses to stay efficient. 

The work spans three layers. The first is the wearable hardware itself, ultrasonic sensing, analog front ends, beamforming ASICs, and power and thermal design, built for continuous clinical use rather than short lab sessions. The second is adaptive decoding: instead of relying on a static decoder, the system uses neuromorphic computing, spiking networks, Bayesian uncertainty estimation, and continual and reinforcement learning to adjust in real time to fatigue, disease progression, and sensor drift, directly on the edge hardware. The third is turning all of this into something that can actually be built and deployed, hardware engineering that meets strict power, thermal, and size constraints while remaining manufacturable and reliable. 

More broadly, I believe the future of intelligence depends not on ever-larger models, but on understanding how biological systems allocate computation under strict energy constraints. The objective is to lay the foundations for energy-efficient intelligent systems that learn continuously, restore neurological function, and expand human capability by combining neuroscience, semiconductor architecture, adaptive machine learning, and first-principles hardware engineering.