What AI, Neuroscience, and Cognitive Science Can Learn from Each Other: An Embedded Perspective

Scientists studying in the fields of AI and neuroscience can learn much from each other, but unfortunately, since about the 1950s, it has been mostly one-sided: neuroscientists have learned from AI, but less so the other way. I argue this is holding back both brain understanding and progress in AI. Current AI (“neural network”/deep learning algorithms) and the brain are very different from each other. The brain does not seem to use trial-and-error–type learning algorithms such as backpropagation to modify weights and more importantly does not require the cumbersome rehearsal needed for trial-and-error implementation. The brain can learn information in a modular and true “one-shot” fashion as the information is encountered while the AI cannot. Instead of backpropagation and rehearsal, there is evidence that the brain regulates its inputs during recognition using regulatory feedback: form the outputs back to inputs—the same inputs that activate the outputs. This is observed through evidence from the fields of neuroscience and cognitive psychology but is not present in current algorithms. Thus, the brain provides an abundance of evidence about its underlying algorithms and while computer science tools and analysis are essential, algorithms guided by computer science should not be standardized into neuroscience theories.