Q&A with Nikolaus Kriegeskorte: Using Artificial Intelligence to Understand the Brain
Computers grow smarter each day, but have yet to reach a human level of intelligence. Biological and digital brains differ in intriguing ways, and closing that gap has become an area of interest for researchers studying the brain and trying to build smarter machines.
Nikolaus Kriegeskorte recently joined Columbia University as a Professor of Psychology and Director of Cognitive Imaging at the Mortimer B. Zuckerman Mind Brain Behavior Institute. A computational neuroscientist, Kriegeskorte uses machine learning algorithms to model how the brain sees and interprets the world. He also helped organize a three-day conference at Columbia, Cognitive Computational Neuroscience (CCN), that starts Sept. 6 and brings together cognitive scientists, neuroscientists, and computer scientists, to share their perspectives on intelligence and the brain. Kriegeskorte spoke with Columbia News about the event and his research.
Q. What inspired this conference?
A. To understand cognition — the brain’s most complex abilities — we need to build models of the brain that can perform feats of intelligence. We can only pull this off if the fields of cognitive science, computational neuroscience, and artificial intelligence work together. That’s the point of this conference. Neuroscientists Tom Naselaris and Kendrick Kay approached me with the idea two years ago and it took only a few minutes to convince me.
Q. How did you get Facebook’s Director of Artificial Intelligence Research, Yann LeCun, and other high-profile researchers to speak?
A. Cognitive science and AI share the goal of modeling intelligent behavior, but until recently haven’t focused much about the brain. At the same time, brain-inspired deep neural network models have led to major breakthroughs in AI. This reinforces the idea that we can understand the mind by looking at the brain. The pieces of the mind-brain puzzle are on the table. Now we have to put them together.
Q. What’s the goal of the conference?
A. I hope it will spark transdisciplinary collaborations and become an annual event that helps to establish cognitive computational neuroscience as a new field. A key challenge is to translate between the languages of different disciplines, and between cognitive models, neural network models, and biologically detailed neuronal-circuit models.
Q. How do neuroscientists get past theories and make discoveries?
A. Historically, microscopes, telescopes and new tools for making observations have led to major discoveries. But better observations do not necessarily lead to discovery. Galileo had Copernicus’s theory to guide his interpretation. Neuroscientists have traditionally gone from single-neuron observations to emergent cognitive functions. But to understand what the neurons are doing and their role in cognition, we also have to think top-down, from computational theories of brain function to the neurobiological details. This is how cognitive science and AI can complement neuroscience.
Q. How are new theories about the brain tested?
A. Let’s say a theory claims to explain how a cognitive function works. We can embed the theory in a computer-simulated model and test whether the model can perform that function. We can further test the model by predicting specific patterns of neuronal activity and behavior. For example, the model should be able to predict when the brain makes mistakes, and what kinds of problems require more time to solve.
Q. How are neuroscientists using machine learning and big data sets to study the brain?
A. The human brain is complex. It has to be to model our environment. This means that cognitive and brain scientists need highly detailed observations. Machine learning can help discover structure in these big data sets. But beyond data analysis, machine learning provides the theoretical foundation and software tools to model the brain. After all, the brain itself is a learning machine.
Q. How are computer scientists using neuroscience to develop better AI since I gather that this is a two-way conversation?
A. The brain has inspired neural network models that can learn millions of connection weights through experience, like biological neural networks. Neural network models now underlie many AI applications. They can take a photograph and recognize the people and objects pictured. They are also good at translating languages and controlling robots. Still, biological brains are far more complicated. Incorporating more of the neurobiological mechanisms into models might lead to further advances in AI.
Q. Is AI reaching human-level intelligence?
A. AI is making rapid progress, but is not there quite yet. Imagine a child seeing an escalator for the first time. She might recognize that it’s like a moving staircase, infer its use, and imagine riding on it. After one experience, and before even learning the term “escalator,” she might form a new concept and quickly recognize it later. Today’s AI systems are still a little dim by comparison.
Q. How do you use machine learning in your research on vision?
A. Our goal is to build better models of human visual cognition. We study how people see and understand the scenes around us, recognizing people and objects and their relationships. If we fully understood the process, we should be able to build a model that can perform these feats of intelligence. In my lab, we build neural network models and test them by comparing their internal representations of images and video with those in biological brains. We measure brain representations with functional magnetic resonance imaging and electrophysiological techniques. We also look at behavior: the patterns of reaction times and errors in various tasks. Whereas engineers are trying to optimize performance, our aim is to model biological brains. So we want our models not only to shine, but also to stumble and fail in ways similar to human brains.
Q. What attracted you to Columbia?
A. The Zuckerman Institute was irresistible to me. On the one hand, it’s an amazing place for experimental neuroscience. On the other hand, it is also world-leading in theoretical neuroscience. My work tries to link these approaches. By taking this position, I hope to be challenged and to learn from my colleagues.
Further reading: CCN blog