Neuroscience and AI: Powerful allies in understanding intelligence

“Our intelligence is what makes us human, and AI is an extension of that quality.” — Yann LeCun

Cutting-edge research findings from neuroscience and artificial intelligence (AI) are coming together to make great strides in our understanding of intelligence a reality. Recent ground-breaking methodological advances in neuroscience now make it possible to collect data from bigger populations of neurons (information-transmitting brain cells) than ever before. But making sense of the sheer quantity of information produced by such “big data” approaches would not be possible without the help of fast and efficient AI algorithms.

One remarkable example of such an approach is the use of selective-plane illumination microscopy (SPIM) to image activity from the whole brain of the juvenile form of a tiny transparent fish, the zebrafish. Researchers modify neurons in the fish to carry a gene which signals brain activity with firework-like flashes of fluorescent light, allowing SPIM to directly measure the activity of nearly every cell in the brain simultaneously (roughly 100,000 neurons in the juvenile fish). This approach has already dramatically improved our understanding of how networks of cells coordinate at a whole-brain level to drive behaviours such as movement and predator avoidance.

Given the huge amounts of image data generated by this technique, machine learning approaches have been critical to allowing efficient processing and labelling of the data, helping to identify active cells, estimate activity rates and infer causal connections. Such impressive developments pave the way to rapidly improving our understanding of the fundamental mechanisms of cognitive systems, yielding significant advances in our understanding of how the brain works.

Fluorescently-labelled cells in the zebrafish brain are helping to illuminate the mysteries of intelligence. (Credit: Marina Venero Galanternik, Daniel Castranova, Tuyet Nguyen, and Brant M. Weinstein, NICHD, NIH)

In turn, insights gained from neuroscience are being applied to develop and refine algorithms for machine learning. Reinforcement learning algorithms, for instance, which tackle the problem of how to act intelligently based on often unpredictable feedback from the world (and have been used to build bots that learn how to play Doom using pixels alone), owe their existence to behavioural learning models developed originally from neuroscience. These algorithms are being further advanced as the basal ganglia, the specialised “feedback-learning” centre of the brain, is better understood.

Likewise, researchers exploring how robots can learn to move and interact fluidly with the world have drawn heavily from models of cognitive development in infants. These build on the key insight from the cognitive sciences that learning how to move in the real world is a complex process which requires developing knowledge of both the environment and one’s own body. Through “learning by doing”, like a toddler would, the robots are saved of the need to be pre-programmed to handle the immense diversity and complexity of situations they are likely to encounter.

Another powerful example is that of recurrent neural networks, machine learning systems formed from looped networks of “neural” units. They are used widely as a kind of “memory” for applications such as sentence comprehension (or even translation), in which the meanings of words read one-by-one must be stored to understand the phrase as a whole. The strikingly similar recurrent connections seen in the hippocampus, an area of the brain crucial for memory, are encouraging dialogue between neuroscientists and machine learning researchers about how hippocampal models might inform recurrent neural network design in future.

Such cross-disciplinary research demonstrates the great benefits that can come from interaction between these two fields. As the level of collaboration increases, a “virtuous circle” can be established in which advances in one field spur on further progress in the other. Better computer vision algorithms help us to understand more about how fluorescent fish see and move, in turn inspiring algorithms for the next generation of autonomous agents (including, ironically, a robotic fish which can fool real zebrafishes into being its friends). Hence, neuroscience and artificial intelligence become powerful allies in the quest to understand, and indeed to create, intelligence. Together, they promise to open our eyes to yet more incredible discoveries and advances over the coming years and decades.