This year, artificial intelligence discoveries were awarded two Nobel prizes. Three of the joint winners have connections with the Gatsby Computational Neuroscience Unit at UCL. Here we look at how their transformative work continues to impact our research.
"Artificial neural networks were initially inspired by the study of biological brains. It is beautiful to see how this has now come full circle, as we use these networks to help us understand the biology of intelligent behaviour in humans and animals," commented Professor Tom Mrsic-Flogel, Director of Sainsbury Wellcome Centre at UCL (SWC), that houses the Gatsby Unit.
"This increased understanding is having a profound impact not only on neuroscience but also on society as AI touches many aspects of our everyday lives," he added.
Transformative impact of Hopfield networks
A member of the founding advisory board of the Gatsby Unit, Professor John Hopfield was awarded one half of the 2024 Nobel Prize in Physics for foundational discoveries and inventions that enable machine learning with artificial neural networks.
Professor Hopfield's pioneering work in the 1980s laid the foundation for modern neural networks. Hopfield networks are characterised by stable states, or attractors, where systems can settle into specific patterns of activity. This concept, originally from physics, has been used to think about how neurons in the brain might learn and store information.
At SWC, Hopfield models have been instrumental in advancing our understanding of how the brain drives behaviour.
Dr Athena Akrami, a Group Leader at SWC, highlights how Hopfield networks have given us a lot of intuitions about memory consolidation and retrieval in the brain: "Hopfield's work has provided a mathematical framework that allows us to think about complex systems at a macroscopic level. This framework has been crucial in shaping our understanding of how the brain makes decisions and forms short-term and long-term memories. We now think of these processes in terms of dynamics inspired by the Hopfield network."
Traditionally, neuroscientists viewed the brain as a series of reflexive, input-output mechanisms, heavily influenced by neurophysiologists like Sherrington and Pavlov. However, Hopfield networks shifted this perspective by emphasising the brain's internal dynamics.
Dr Jeffrey Erlich, Group Leader at SWC, explained: "Instead of thinking of the brain as an input-output machine, we can imagine it more like a body of water that is constantly moving. Whenever there is a stimulus, it changes the waves that are generated, thereby modifying the ongoing system. In the absence of all input, it does not decay to zero as it has its own internal dynamics,"
This concept of internal dynamics has shifted the focus from reactive models to understanding the brain's ongoing activity and stability. For instance, these networks can perform pattern completion, recognising and completing partial or noisy inputs based on learned patterns.
Hinton: the Godfather of AI
Recognising patterns has also been key to another impact of the Nobel discoveries on neuroscience. The founding director of the Gatsby Unit, Professor Geoffrey Hinton, known as the "Godfather of AI," was awarded the other half of the 2024 Nobel Prize in Physics. Professor Hinton is particularly revered for his pioneering work on deep learning.
Dr Adam Tyson, Head Research Engineer at SWC, added: "Deep learning has transformed so many areas of science and society, from detecting cancers on MRI scans to allowing us to use virtual backgrounds on video calls. Much of the data analysis we do at SWC relies on deep learning, particularly for analysing images and videos. In my team we also build tools that use deep learning, for example cellfinder is a software for automated 3D cell detection in very large 3D images."
Deep learning has also had a wider impact on the field of neuroscience. Because deep learning requires significant computation, manufacturers began adapting and making graphics processing units (GPUs) to allow people to run machine learning workflows on their computers. This led to the hardware becoming more readily available, allowing scientists to use these GPUs to accelerate data analysis.
This ability to accelerate data analysis is allowing scientists at SWC to pursue bolder research questions. For example, only a few years ago video data from experiments needed to be manually scored. Now we have tools that can automate this process with very high accuracy, allowing researchers to study more naturalistic behaviour over longer periods of time.
In addition to the positive benefits, Professor Hinton is also known for speaking out about the potential dangers of AI, a concern that SWC takes very seriously.
Dr Tyson explained: "While our work is in a relatively safe space, we think a lot about making sure everything is validated. Deep learning networks can hallucinate outputs, and they can often produce plausible but not necessarily correct results. And so, we make sure everything is objectively tested and validated, as our aim is to build tools that we trust and that will work well for labs around the world."
Hassabis: using AI to solve complex problems
AlphaFold2 is a prime example of an AI tool that has helped many people. Developed by UCL Queen Square Institute of Neurology alumnus Sir Demis Hassabis and the team at Google DeepMind, the tool won Sir Demis and colleague John Jumper one half of the Nobel Prize in Chemistry 2024 for protein structure prediction. Sir Demis also completed at post-doc at the Gatsby Unit before co-founding Google DeepMind.
Clémentine Dominé, a PhD student at the Gatsby Unit who used predictions from AlphaFold2 during her rotation, highlighted how transformative the tool has been: "AlphaFold allows us to predict the 3D structure of proteins from their sequences, which was previously a very expensive and time-consuming process involving crystallography. This has significant implications for understanding protein interactions and developing new therapies. My PhD rotation focused on whether access to this 3D information can improve models to predict the binding between antibody and antigens, and thereby help create better vaccines."
The tool is built on an algorithm called Deep Q-Learning, which allows it to learn policies on complex problems. Scientists at SWC are also using Deep Q network (DQN) models in their research. For example, the Erlich and Duan labs are using DQN in their research on multi-agent decision-making.
Collaborative spirit of SWC and the Gatsby Unit
These advances in AI continue to help us pursue bold questions that were previously not possible.
Professor Maneesh Sahani, Director of the Gatsby Computational Neuroscience Unit at UCL, commented: "AI is having an undeniable impact on science and society. Machine learning is revolutionising fields and propelling academic disciplines forward. We see this at the Gatsby Unit where, besides continuing to push forward the foundational mathematics of ML, we combine with SWC to bring a critical mass of theoreticians and experimentalists together to solve complex problems."
The interdisciplinary culture at SWC and the Gatsby Unit, coupled with the scientific freedom provided by core funding from the Gatsby Charitable Foundation and Wellcome, continues to drive forward our understanding of intelligence in brains and machines, offering exciting possibilities for the future.
This article was written by April Cashin-Garbutt and originally published by the Sainsbury Wellcome Centre at UCL.
