NEW YORK – These are thrilling times to be a neuroscientist. The same goes if you’re on the forefront of artificial intelligence. So what is it like to be part of a hybrid research frontier, NeuroAI, with ambitions to take both AI and brain science to new heights? To find out, we sought perspectives from a few of the many people at Columbia who are revving up momentum in NeuroAI:
Kenneth Miller, PhD, a cofounder of the Zuckerman Institute’s Center for Theoretical Neuroscience, has spent his career working to understand the cerebral cortex and the nature of its biological computations.
Richard Zemel, PhD, director of Columbia’s new AI Institute for Artificial and Natural Intelligence (ARNI), supported by the National Science Foundation, is a computer scientist and engineer and a developer of advanced artificial intelligence systems.
Nikolaus Kriegeskorte, PhD, is a computational neuroscientist, builder of computer models that help reveal how our brains enable us to see and a co-founder of the Cognitive Computational Neuroscience conference.
What is NeuroAI?
Dr. Kriegeskorte: NeuroAI is an emerging field between neuroscience and artificial intelligence (AI) engineering. Computational insights from AI help us understand how our brains process information. Conversely, brains still beat AI in many ways, so engineers can learn from biology.
Dr. Zemel: Neuroscientists seek explanations for how the brain carries out a variety of perceptual, cognitive and motor tasks. We think that AI can provide insights here.
Dr. Miller: Modern AI systems are based on the idea that computations in brains occur without explicit logic or symbols, through neurons summing inputs from other neurons, and that learning occurs by resetting synaptic strengths between neurons to specify how much each neuron influences those it connects to. We can train these neural networks to perform in computers the same tasks done by parts of the brain, which means their computational traits can lead to hypotheses about how brains compute. AI also is providing new methods for analyzing brain and behavioral data, helping us decode the meaning of brain activity.
Dr. Zemel: At the same time, what we are learning in neuroscience can also improve AI systems.
Dr. Kriegeskorte: Current AI systems are good at one thing, like identifying what’s in a picture, but they don’t generalize as robustly as our brains to new challenges, like recognizing the same object in an unexpected context. NeuroAI seeks lessons in biology to build AI systems that are as versatile and efficient as our brains.
Why is NeuroAI emerging now?
Dr. Kriegeskorte: About a decade ago, broadly brain-inspired neural network models began to deliver on their promises. Neural network models have since revolutionized AI. In parallel, the modern neuroscience revolution gave us much richer ways to measure and manipulate brain activity. These two complementary revolutions have disrupted both fields. Some researchers in each field are looking to the other field for guidance.
Dr. Zemel: Major technical advances have been revolutionizing our ability to observe and manipulate brains at a large-scale and quantify complex behaviors. Recent AI systems have also made remarkable progress in solving tasks, often at superhuman levels of accuracy. The latest AI systems have surprised even the researchers in the field with their abilities, such as adapting to new domains and engaging in reasonable dialogues.
Dr. Miller: A big breakthrough occurred in 2012, when deep neural networks were shown for the first time to greatly outperform older approaches to artificial intelligence. This was made possible by the availability of big computational power and big labeled datasets used to train AI.
Dr. Zemel: Lagging is our understanding of natural brains and intelligence as well as what makes our new AI systems so powerful. We think it will take a new field like NeuroAI to fill these knowledge gaps.
Dr. Kriegeskorte: With this in mind, community-building meetings, such as the Cognitive Computational Neuroscience conference, have started to bring together the neuroscience and AI communities. Professor Xaq Pitkow of Carnegie Mellon University and collaborators, including some of us at Columbia, developed a NeuroAI course offered this summer by the NeuroMatch Academy to students around the world. We think it will help define this emerging field.
How is Columbia involved in NeuroAI?
Dr. Kriegeskorte: Columbia’s Center for Theoretical Neuroscience is absolutely key to all of this. Housed at the Zuckerman Institute, it brings together a unique community that links neuroscience to statistics, mathematical modeling and machine learning.
Dr. Zemel: Columbia has one of the world’s strongest groups of researchers in machine learning, and theoretical and cognitive neuroscience. They, along with researchers from 12 other institutions, also form the core of our new AI Institute for Artificial and Natural Intelligence, or ARNI. Led by myself, ARNI aims to foster the cross-disciplinary research that underlies NeuroAI. Reflecting that commitment are the contributions that Dr. Pitkow, ARNI’s associate director, and many of our ARNI colleagues made to the NeuroAI course Dr. Kriegeskorte mentioned.
Dr. Kriegeskorte: ARNI is a major inspiration because it brings me together with engineers and biologists who understand things I need to learn for my own research in computational neuroscience.
Dr. Miller: I’ll add that ARNI is opening pathways to the field’s future by funding postdoctoral fellows who will build the bridges between those of us who until now have been focusing either on natural or artificial intelligence alone, but who now are finding common ground in NeuroAI.
How do you envision NeuroAI unfolding in the coming years?
Dr. Miller: People will keep developing creative AI solutions to computing and learning that provide models and hypotheses about brain function. Learning in AI systems often uses algorithms or implementations that brains cannot, so neuroscientists will continue uncovering the brain’s methods for solving the same problems. NeuroAI will attempt to improve the performance of AIs by better emulating properties of neurons and brains. AI is already helping us decode brain activity and improve brain-computer interfaces that allow the disabled to write, speak and control prosthetic limbs and hands. It’s also upping our power to analyze neural data. Advances in this synergy of neuroscience and AI are so rapid now that we cannot predict all the ways NeuroAI will change our lives.
Dr. Zemel: The hypothesis is that NeuroAI will ultimately lead to new neuroscientific theories and, on the AI side, to advanced datasets and more efficient computing frameworks that will lead to AI capable of novel learning and reasoning tasks.
Dr. Kriegeskorte: Making a generalized AI with the versatility of our own brains is a famously hard problem. That’s because there’s no right way to generalize. Different AI models embody different biases and different prior assumptions. An AI that works well in some domains, like in writing tasks or recognizing objects, will likely not work well in others. The key question is whether NeuroAI will deliver AI that, like the human brain, can do it all.
###
More about our faculty:
Kenneth Miller, PhD: Peter Taylor professor of neuroscience at Columbia’s Vagelos College of Physicians and Surgeons; principal investigator at Columbia's Zuckerman Institute, member of the Executive Committee of the NSF AI Institute for Artificial and Natural Intelligence and co-director of Columbia’s Center for Theoretical Neuroscience
Richard Zemel, PhD: Trianthe Dakolias Professor of Engineering and Applied Science, professor of computer science and director of the NSF AI Institute for Artificial and Natural Intelligence.
Nikolaus Kriegeskorte, PhD: professor of psychology, professor of neuroscience at Columbia’s Vagelos College of Physicians and Surgeons, principal investigator at Columbia's Zuckerman Institute, academic collaborator with the NSF AI Institute for Artificial and Natural Intelligence.
