Margot Webster investigates the relationship between neuroscience and artificial intelligence.
Written by: Margot Webster
Art by: Sotiria Kal
Next year will mark the twentieth anniversary of the Gatsby Computational Neuroscience Unit at UCL, which was created in an effort to promote collaboration between experts in machine learning and theoretical neuroscientists. This endeavour has certainly proved its worth, as many of the unit’s alumni have gone on to lead the field of Artificial Intelligence (AI), most notably Demis Hassabis, who founded and runs Google’s DeepMind.
I met with Professor Michael Häusser at the Wolfson Institute for Biomedical Research, where he runs a laboratory focused on neural computation. He describes his work as trying to unravel ‘the code that the brain uses to process and store information’. AI developers are perhaps doing the inverse, by trying to invent the codes and algorithms able to support complex functions. The computing device inside our heads, capable of performing most of these functions, seems like a logical source of inspiration for AI developers. In fact, there is currently ‘an explosion of this kind of collaboration’, according to Häusser.
Understanding the processes of the human brain has proved fruitful for artificial intelligence in the past. The hierarchical architecture of the visual system has been successfully reproduced in artificial neural networks that can recognise faces in photos (think FaceID on the new iPhone). The way that humans and animals make associations in the hope of maximising rewards, called ‘reinforcement learning’, has also been implemented in AI. These methods enable predictions to be made about which decisions lead to rewards, such as winning a game of backgammon.
Although machines can now outperform humans at a number of different tasks, from video games to chess, the human brain’s ability to interact with the real world, learn from a single experience, and enable creativity has yet to be matched. The human brain is highly complex and unlike the building blocks currently used in AI, real neurons are astoundingly diverse and can each perform complex computations. However, ‘it’s not clear to what extent this diversity is directly relevant to the computations that the individual neurons are performing’, says Häusser. This could explain why such diversity hasn’t been exploited in AI yet. Evolution is essentially a random process, so observing certain characteristics in the human brain does not mean that these are necessary to perform certain functions, and that they can, or should be reproduced in machines.
One aspect of the human brain that is far from being matched by AI at present is its impressive energy efficiency. This is an important factor for applications of AI in a world that emphasises portability and small size above all else. The brain can perform complex tasks using very little energy (about as much as a light bulb!), and is ‘orders of magnitude more efficient than the equivalent computer you would need to achieve the same level of performance’, Häusser says. ‘There are real lessons to be learned from the brain about how to implement energy-efficient computations’.
In return, neuroscientists can use new technologies exploiting AI techniques to process and analyse data faster and more efficiently than ever before. However, there may be more to this relationship than the exchange of tools for knowledge of brain mechanisms. Indeed, ‘the most fruitful interface might be areas where there isn’t that kind of dichotomy’, says Häusser. ‘Machine learning approaches may reveal brain mechanisms which in turn inspire new approaches that can be exploited for AI’. It is also conceivable that AI might implement algorithms that can then be used by neuroscientists as a framework for understanding data recorded from the brain.
Neuroscience and technology have been merged in the past, with the development of Brain-Machine Interfaces (BMIs). Robotic arms can be controlled by a person’s thoughts, via electrodes that read out the activity of the motor cortex (the part of the brain that controls movement) and translate them into commands for an artificial limb. With AI technology thriving, the development of these interfaces shows no sign of slowing down. Companies such as Elon Musk’s start-up Neuralink, launched earlier this year, are hoping to wire brains directly to intelligent machines. As Häusser told me at the end of our interview, ‘it’s a tremendously exciting time […] but also a time to reflect on potential ethical implications’.