Training a single large AI model can now emit as much carbon as three round-trip flights between New York and San Francisco. This fellowship tackles that waste by building a fundamentally different kind of AI—one inspired by the human brain, which makes complex decisions on just 20 Watts of power. The problem is twofold. Today’s AI runs on conventional computer chips that physically separate memory from processing, forcing data to shuttle back and forth at high precision. The brain, by contrast, uses noisy electrical spikes and merges memory with computation. This project draws on the free energy principle, a theory that explains how brains minimise surprise using inherently probabilistic models. The key insight: those models do not need high-precision data, so the “imperfections” of nanoscale devices—normally a drawback—can be harnessed as a computational resource. If successful, the research could deliver AI hardware that is 1,000 times more efficient than current systems, running on drastically lower voltages and currents. That would shrink the carbon footprint of everything from language translation to protein-folding models, and enable embedded AI in devices too power-constrained for today’s chips. The work is fundamental science, but it directly targets a practical bottleneck that already shapes the cost and environmental impact of modern computing.
View original technical description
This fellowship will lay the foundations for a new AI paradigm featuring algorithms based on the free energy principle (FEP) and hardware platforms leveraging the stochasticity of novel nanoscale devices based on 2-dimensional materials, enabling embedded systems with unprecedented efficiency. Artificial Intelligence (AI) models based on deep learning algorithms have demonstrated super-human performance for a wide variety of such tasks - ranging from language translation to protein folding. However, the cost of developing such models - both in terms of energy and time - has been sky-rocketing. For example, recent studies estimate that the carbon footprint for training a state-of-the-art language translation model can be as high as 3 round-trip flights between New York and San Francisco. One major contributor to this inefficiency is the von Neumann architecture used in today's computing platforms - the data storage and data processing units are physically separated. Hence, running these algorithms require data that is represented in high precision to be constantly shuttled back and forth. Contrast this with the human brain, nature's most evolved computation engine, which continuously makes complex cognitive decisions, that too based on noisy sensory data and an imprecise computational infrastructure. The brain achieves this amazing feat by encoding information in tiny electrical signals called spikes that are transmitted through a seamlessly interconnected network of 'logic' and 'memory' units - neurons and synapses - all while consuming less than 20 Watts. Clearly, there is something fundamentally unique about the algorithms and hardware of the brain! The research in this fellowship is motivated by a theory called the free energy principle (FEP), which provides a unified foundation that underlies the cognitive efficiency of the brain. The central tenet of FEP is that biological organisms tend to minimize the occurrence of surprising events by acting to change the sensory inputs they receive from the environment or by modifying the internal states that allow them to perceive the world and make decisions. Furthermore, since the theoretical foundation of FEP assumes that the brain's models are inherently probabilistic, representing data or the model in high precision is not a strict requirement. Hence, the research in the fellowship will pursue the novel approach of using the undesirable imperfections of nanoscale devices as a resource for implementing the probabilistic parameters of the model. This approach can hence lead to computational systems with unprecedented efficiency as the basic building blocks can be operated at drastically lower voltages and currents, avoiding unnecessary data movement. This research will first develop artificial neural networks that mimic the spike-triggered communication feature of the brain based on the mathematical ideas of the free energy principle. We will create AI models that can generate decisions that are trustworthy and can be supported with quantifiable confidence metrics. In parallel, we will also demonstrate prototype hardware platforms that implement these algorithms using the stochastic properties of nanoscale devices as a resource for computation. Hardware prototypes will be built using novel nanoscale devices that are based on 2-dimensional materials as well as nanoscale memory arrays built by industrial partners targeting a 1000-fold improvement in computational efficiency compared to what is possible today. Thus, the fellowship will lay the foundations of a new Smart and Green AI paradigm.
Plain English summaries and category classifications on this site are generated by AI and may not perfectly reflect the original research.
Is something wrong? Let us know