Traditional notions of artificial intelligence (AI) are outdated. Rather than simply being able to understand inventory lists, or make binary decisions based on prescriptive rules, the field now trying to mechanically replicate human intelligence. And to do so, it requires a fundamental understanding of how the brain works- something increasingly found in neuroscience.
In the last decade, insights from neuroscience have had a large influence in propelling AI. Artificial neural networks for example were inspired by the neural networks found in the brain, and have given computers the ability to distinguish between cats and trees, as well as how to spot pedestrians well enough to avoid crashing into when used by self-driving cars. Meanwhile, neuroscience has also begun to benefit from AI. Thanks to recent advances in the field, neuroscientists have been able to both develop and test ideas about how the brain computes information, as well as process complex data sets.
But even though interactions between the two disciplines have mutually propelled each other forwards, some problems remain to be solved. Although machine learning, the technology behind artificial neural networks, is more efficient than humans at finding complex and subtle patterns in very large data sets, it still pales in comparison to the human brain in other tasks. Working by analyzing enormous amounts of data to recognize patterns- whether that’s to be able to differentiate between faces, or know when to take a right turn on a specific road- unlike humans, it is unable to make swift conclusions from smaller datasets, and then extrapolate these and reapply them in novel contexts.
For example, while a 3 year old child is able to correctly deploy grammatical rules in their mother tongue on a limited number of inputs- or even more simply, recognize objects- an AI currently needs exponentially more data to yield a similar result. Why? The answer is simple. Although human neural networks have inspired current forms of AI, our incomplete understanding of how human intelligence operates means that AI still works very differently from the human brain.
While an AI may require millions of datasets to be able to correctly identify a cat, after being shown a singular cat photograph, a 3 year old child is not only be able to recognise cats therefrom, but also cats in previously unseen images, sporting different colors, posing at different angles and represented in more abstract ways such as line drawings or pixelated mosaics. Although given enough datasets an AI may be able to replicate these results to some degree, unlike a 3 year old human, off of a single data set, it would remain at a loss.
To solve this issue, researchers are now investigating ways that enable neural networks to perform “unsupervised learning” ie. learning that does not involve enormous quantities of labeled datasets. Daniel Yamins for example, a computational neuroscientist at the Wu Tsai Neurosciences Institute at Stanford University in California, is currently devising programs that replicate babies how babies learn. Integrating their environment via random interactions that slowly develop frameworks of how the world works, he essentially codes curiosity into each program in the hope that it will motivate it enough to explore, and thus create new behaviors.
Another biological avenue for AI research comes from the notion that aspects of our intelligence may have been installed over time by evolution. For example, unlike an AI that would need to be exposed to millions of images of faces to be able to recognise a face as a face, babies are able to do so from their first hours of life. To this, Tomaso Poggio, a computational neuroscientists at the Center for brains, Minds and Machines at MIT suggests that our genes may in fact encode mechanisms for quick and early learning. Understanding to what extent this is the case may provide a way to better replicate human intelligence in machines.
While the goal of neuroscience is to ultimately understand how the human brain works, Yamins says that it will be unable to do so without collaboration with AI. He said, “If you don’t have an AI solution, if you have nothing that works artificially, you can’t possibly have a model of the brain.” Ultimately, as the symbiosis between AI and neuroscience continues to develop, it is likely that computer scientists will continue to leverage neuroscience to create better models of the human brain. To understand how accurately they represent human intelligence however, it will likely be neuroscientists who test them out, and then have the final say.