The fields of artificial intelligence (AI) and neuroscience are closely intertwined. Artificial intelligence was inspired by the human brain, and, in turn, AI can help us better understand the brain's complex inner workings.
AI's ability to make sense of huge datasets is enabling new developments in neuroscience technology and health care. At the same time, AI models offer a powerful way to study how we learn, perceive, and even feel emotions.
Help neuroscientists make sense of large, complex datasets.
Enable brain–machine interfaces that assist patients.
Create models of the brain to understand its inner workings.
Finding Insights in Big Data
Our understanding of the brain is shaped by data—and lots of it. Inside the human brain, tens of billions of neurons communicate with one another trillions of times a second. Neuroscientists record this activity with special tools that measure electrical and chemical signals in the brain. This results in far too much data for humans alone to sift through and analyze efficiently.
AI tools, however, can quickly analyze large amounts of data. One of the greatest strengths of AI is its ability to reveal patterns in data that are too subtle for humans to spot. When applied to neural data, AI tools can extrapolate patterns across large datasets of brain activity.
In one study, a machine learning algorithm was trained to analyze brain scans and detect Alzheimer's disease with a 90 percent accuracy rate. The same pattern-recognition capabilities can be applied to study the brain using genetic or video data. For example, Caltech scientists Pietro Perona, the Allen E. Puckett Professor of Electrical Engineering; and David Anderson, the Seymour Benzer Professor of Biology, director and leadership chair of the Tianqiao and Chrissy Chen Institute for Neuroscience at Caltech, and Howard Hughes Medical Institute Investigator; and members of their research groups have created AI vision tools to analyze videos of animal behavior recorded during experiments. The tools can automatically detect and make note of shifts in position, posture, or activity that are hard to spot with the naked eye. These tools also work far more quickly than humans making the same annotations by hand, allowing scientists to process more data and free up time to interpret results.
Brain–Machine Interfaces
Other sources of big data that benefit from AI-powered analysis are brain–machine interfaces (BMIs). In a BMI system, electrodes are implanted in a patient's brain, and the brain signals are processed and decoded in real time to communicate with a computer or to control a prosthetic device such as a robotic hand. Because the amount of data coming from the brain is so large, AI can help process data quickly and detect meaningful brain activity.
At Caltech, Richard Andersen, the James G. Boswell Professor of Neuroscience and director and leadership chair of the T&C Chen Brain-Machine Interface Center, and his research group have created a BMI that allows paralyzed patients to move a prosthetic arm or a cursor on a computer screen with only their thoughts. Caltech's Azita Emami, the Andrew and Peggy Cherng Professor of Electrical Engineering and Medical Engineering, and director of the Center for Sensing to Intelligence, collaborates with Andersen to create the hardware and AI to power BMIs. Using machine learning algorithms, Emami's group even created an implant that can predict when a seizure is coming for patients with epilepsy. The use of AI enables more personalized data analysis, and, therefore, more accurate results for each patient.
Modeling the Brain
The brain is often referred to as a black box—a system whose internal workings are largely a mystery. Even neuroscientists acknowledge that we still understand very little about how the brain actually works. One approach to solving the brain's mysteries is to create a representative model and examine how the model works.
Computational neuroscientists create models of the brain by using mathematical equations to create a simplified representation of brain activity. The simpler the model, the easier it is to understand. Of course, the brain is anything but simple, so computational models aim to strike a balance between accuracy and clarity.
Researchers train machine learning algorithms on recorded brain data until the algorithms can produce similar results to the real data. Although there is no guarantee that these models work exactly the same way a brain does, scientists can examine the models, compare them to our theories about the brain, and make predictions about what data a real experiment will produce based on the model's outputs.
"Trying to build intelligent systems makes us think harder about the structure of such systems, useful representations, and algorithms, and, thus, it offers us hypotheses about how the brain may be structured and what different signals in the brain may represent," Perona says. "Also, trying to build intelligent systems makes us think harder about what the machine is trying to accomplish and how to measure its performance—this is useful to neuroscientists because it helps them be clear about what questions they are asking."
In a different study from the lab of David Anderson, researchers trained a machine learning model on brain recordings of mice in order to study how the brain feels anger and aggression. The model displayed a pattern of activity that could explain how and why a signal of anger is encoded into neural circuits.
In another example, Caltech neuroscientist John O'Doherty, the Fletcher Jones Professor of Decision Neuroscience, compared the "neural" activity of an AI model that learned to play Atari video games with that of humans playing the same games. O'Doherty's lab found that activity in the model looked similar to the activity in the human brain, leading them to believe the AI model could teach us how humans perceive space and make decisions.
"The interaction between AI and neuroscience goes both ways," O'Doherty says. "If we can find out how similar AI algorithms are to the brain, this helps us better understand how the brain solves these kinds of hard problems, but, conversely, if we can understand why and how the brain can solve these games much more efficiently compared to an AI, this may help guide the development of smarter and more humanlike AI algorithms in the future."
