Neuroscience versus artificial intelligence engineering: We asked a leading exponent in each field to weigh in on where the two meet.

Jonathan Whitlock

Head of the Whitlock research group at Kavli Institute for Systems Neuroscience, Jonathan Whitlock's work is focused on unraveling the neural circuits behind how we plan our own actions and understand the actions of others.

1. Has anything in your neuroscientific research and understanding led you to think that a human mind could be replicated as AI?
   The more I learn about how bewilderingly complicated the brain of a rat is, the more I think we will never be able to physically re-create a piece of hardware as such. That said, the computations performed by various brain networks are by no means sacred, and most certainly will be matched or exceeded by AI. The key here is generality – artificial general intelligence – which is still just an idea on the horizon. If it were possible to make an artificial neural network (ANN) capable of acquiring and learning to solve problems on its own, that would be a historical landmark, and a major step toward a human-like mind. But in my lifetime, I’d be happy to see an AI mouse.
   
   
2. Are emotions an important or vital component in how the mind works?
   Regardless of whether one thinks of emotions as important to how the mind works, the fact is that the mind and emotions co-evolved over countless eons together. There are certainly various neural computations which are invariant to emotional status (e.g., performing algebra), but on the whole the mind works day-to-day with some kind of registered emotional status and our mood can indeed have a huge effect on the way we think on a given day. Emotions may well have evolved as internal extensions of motivational states (fear, pleasure, etc.), which are the primary forces which drive behavior in all living animals.
   
   
3. What AI developments have made you think differently about the brain and its workings?
   I really liked the work of Olshausen and Field in 1996 – they showed the receptive field properties seen in the primary visual cortex could be recapitulated by unsupervised learning algorithms using maximally sparse coding strategies. It made me think of convergent evolution – it was an instance where the visual system and a computer algorithm converged on the same (or very similar) coding principles to encode natural scenes. It makes one realize that it is possible to recapitulate biological coding principles using unsupervised approaches, and it could probably work for other modalities and for higher dimensional representations as well. The only limit is the programmers and the limits of the computers we can use. My own research in behavioral planning and action perception has yet to be impacted, but it would be a great day indeed if AI got to the point where it paralleled these functions of the brain.

Erik Brynjolfsson

Director of the MIT Initiative on the Digital ­Economy, professor at MIT Sloan School and co-author with Andrew McAfee of The Second Machine Age, Erik Brynjolfsson specializes in the effects of information technologies on business strategy, digital commerce and intangible assets.

1. What neuroscientific research or insights have made you think differently about the brain and its workings and how might they be applied to AI development?
   Different parts of the brain are used when addressing ethical questions, like the famous "trolley problem."* Specifically, duty-based moral decisions tend to trigger automatic emotional responses, while more utilitarian judgements are more likely to require conscious reasoning. If we want robots to make decisions that have an ethical component, then we may need to think hard about how these different systems might apply in artificial intelligence.
   
   *A moral dilemma about the choice to divert an imagined runaway railway trolley where both outcomes will end in differing casualties.
2. What developments in AI give you reason to think that AI can and will (or won't) be able to mimic the fundamental features of the human mind?
   Much of AI research implicitly or explicitly seeks to mimic the human mind – after all, that provides "existence proof" that a particular problem or task is at least possible. I think this is laudable and have little doubt that it can ultimately be achieved, if only because the human mind is itself a physical system, made of atoms, and subject to the same laws of physics as any material object.
   
   I encourage researchers to go beyond using the human mind as their only, or even primary, template. Computers have so many different strengths and weaknesses that we can create superhuman intelligence on some dimensions, even as we fall short on others. Ultimately, having distinct strengths and weaknesses compared with humans makes it more likely that humans and machines can complement each other in a way that creates shared prosperity.
   
   
3. Are emotions an important component in AI research?
   It depends on what you mean by emotions. Reinforcement learning depends very much on AI systems seeking to increase rewards and avoid punishments. Perhaps that can be thought of as a primitive kind of emotion. Another area that is increasingly important is sentiment analysis in natural language processing and detecting emotions in facial expressions. In many applications, like call centers, text analysis, voice recognition or medical diagnoses, machines can do a better job if they understand the emotions of humans.

Curious about the contents of our newest Think:Act magazine? Receive your very own copy by signing up now! Subscribe here to receive our Think:Act magazine and the latest news from Roland Berger.