Researchers Chart Roadmap for Brain-Like Computing to Outpace Traditional AI

How interdisciplinary collaboration and brain-inspired designs could slash AI’s energy costs and unlock real-time learning for robots, medical devices, and beyond.

Review: Neuromorphic computing at scale. Image Credit: Owlie Productions / Shutterstock

Neuromorphic computing—a field that applies principles of neuroscience to computing systems to mimic the brain’s function and structure—needs to scale up if it is to compete with current computing methods effectively. In a review published in the journal Nature, 23 researchers, including two from the University of California San Diego, present a detailed roadmap of what needs to happen to reach that goal. The article offers a new and practical perspective toward approaching the cognitive capacity of the human brain with comparable form factor and power consumption.

"We do not anticipate that there will be a one-size-fits-all solution for neuromorphic systems at scale but rather a range of neuromorphic hardware solutions tailored to specific applications," the authors write.

Neuromorphic computing has applications in scientific computing, artificial intelligence, augmented and virtual reality, wearables, smart farming, smart cities, and more. It has the potential to outperform traditional computers in energy and space efficiency and performance. This could present substantial advantages across various domains, including AI, health care, and robotics. Key challenges include hardware/software co-design, standardized benchmarking, and lifelong learning algorithms, which the paper identifies as critical to advancing the field. As AI’s electricity consumption is projected to double by 2026, neuromorphic computing emerges as a promising solution.

"Neuromorphic computing is particularly relevant today when we are witnessing the untenable scaling of power- and resource-hungry AI systems," said Gert Cauwenberghs, a Distinguished Professor in the UC San Diego Shu Chien-Gene Lay Department of Bioengineering and one of the paper’s co-authors.

Neuromorphic computing is at a pivotal moment, said Dhireesha Kudithipudi, the Robert F. McDermott Endowed Chair at the University of Texas San Antonio and the paper’s corresponding author. "We are now at a point where there is a tremendous opportunity to build new architectures and open frameworks that can be deployed in commercial applications," she said. "I strongly believe that fostering tight collaboration between industry and academia is the key to shaping the future of this field. This collaboration is reflected in our team of co-authors."

Last year, Cauwenberghs and Kudithipudi secured a $4 million grant from the National Science Foundation to launch THOR: The Neuromorphic Commons. This first-of-its-kind research network provides access to open neuromorphic computing hardware and tools to support interdisciplinary and collaborative research.

In 2022, a neuromorphic chip designed by a team led by Cauwenberghs showed that these chips could be highly dynamic and versatile without compromising accuracy and efficiency. The NeuRRAM chip runs computations directly in memory. It can run a wide variety of AI applications—all at a fraction of the energy consumed by computing platforms for general-purpose AI computing. The paper highlights similar advances in emerging memory technologies and notes that "our Nature review article offers a perspective on further extensions of neuromorphic AI systems in silicon and emerging chip technologies to approach both the massive scale and the extreme efficiency of self-learning capacity in the mammalian brain," said Cauwenberghs.

To achieve scale in neuromorphic computing, the authors propose several key features that must be optimized, including sparsity, a defining feature of the human brain. The brain develops by forming numerous neural connections (densification) before selectively pruning most of them. This strategy optimizes spatial efficiency while retaining information at high fidelity. If successfully emulated, this feature could enable neuromorphic systems that are significantly more energy-efficient and compact. The paper also stresses the need to address gaps in software toolchains compared to mainstream AI frameworks, which currently limit adoption.

"The expandable scalability and superior efficiency derive from massive parallelism and hierarchical structure in neural representation, combining dense local synaptic connectivity within neurosynaptic cores modeled after the brain’s gray matter with sparse global connectivity in neural communication across cores modeling the brain’s white matter, facilitated through high-bandwidth reconfigurable interconnects on-chip and hierarchically structured interconnects across chips," said Cauwenberghs.

"This publication shows tremendous potential for the use of neuromorphic computing at scale for real-life applications. At the San Diego Supercomputer Center, we bring new computing architectures to the national user community. This collaborative work paves the path for bringing a neuromorphic resource for the national user community," said Amitava Majumdar, director of the division of Data-Enabled Scientific Computing at SDSC here on the UC San Diego campus, and one of the paper’s co-authors.

In addition, the authors call for stronger collaborations within academia and between academia and industry, as well as for the development of a wider array of user-friendly programming languages to lower the barrier to entry into the field. They believe this would foster increased collaboration, particularly across disciplines and industries.