Another key challenge in contemporary AI research is known as transfer learning. To be able to deal effectively with novel situations, artificial agents need the ability to build on existing knowledge to make sensible decisions. Humans are already good at this – an individual who can drive a car, use a laptop or chair a meeting are usually able to cope even when confronted by an unfamiliar vehicle, operating system or social situation.
Researchers are now starting to take the first steps towards understanding how this might be possible in artificial systems. For example, a new class of network architecture known as a “progressive network” can use knowledge learned in one video game to learn another. The same architecture has also been shown to transfer knowledge from a simulated robotic arm to a real-world arm, massively reducing the training time. Intriguingly, these networks bear some similarities to models of sequential task learning in humans. These tantalising links suggest that there are great opportunities for future AI research to learn from work in neuroscience.
But this exchange of knowledge cannot be a one-way street. Neuroscience can also benefit from AI research. Take the idea of reinforcement learning – one of the central approaches in contemporary AI research. Although the original idea came from theories of animal learning in psychology, it was developed and elaborated by machine learning researchers. These later ideas fed back into neuroscience to help us understand neurophysiological phenomena, such as the firing properties of dopamine neurons in the mammalian basal ganglia.
This back and forth is essential if both fields are to continue to build on each other’s insights, creating a virtuous circle whereby AI researchers use ideas from neuroscience to build new technology, and neuroscientists learn from the behaviour of artificial agents to better interpret biological brains. Indeed, this cycle will likely accelerate thanks to recent advances, such as optogenetics, that allow us to precisely measure and manipulate brain activity, yielding vast quantities of data that can be analysed with tools from machine learning.
We therefore believe distilling intelligence into algorithms and comparing them to the human brain is now vital. Not only could it bolster our quest to develop AI, a tool that we hope will create new knowledge and push forward scientific discovery, but may also allow us to better understand what’s going on inside our own heads. That could shed light on some of the most enduring mysteries in neuroscience, such as the nature of creativity, dreams and, perhaps one day, even consciousness. With so much at stake, the need for the field of neuroscience and AI to come together is now more urgent than ever before.
Download the full paper here.