Caltech: The Caltech Effect February 2021 – Neural Networking


June 25, 2021

The Tianqiao and Chrissy Chen Institute for Neuroscience officially moved into its new building on the Caltech campus in January 2021. Since the institute was inaugurated in 2016, however, researchers from a wide variety of disciplines and interests have been affiliated with Chen, including these five graduate students.

A fruit fly in flight is bombarded by stimuli, from other flies and obstacles it must avoid to signals that will help it to track down a food source. How does an insect with such a small brain adjust instantaneously? Annie Erickson, a Chen Graduate Fellow in the lab of Michael Dickinson, Esther M. and Abe M. Zarem Professor of Bioengineering and Aeronautics, studies the pathways in the fruit fly brain when in flight. “There is a whole class of neurons that project from the brain down to thoracic regions,” Erickson says. These descending neurons get information from sensory areas of the brain, connect with the fly brain’s navigational centers, and send instructions down to motor centers. But the exact ways in which so many different parts of the fly brain collaborate has not been fully understood and mapped. “We’re looking into all these different types of neurons and how they work together.”

Because it would not be practical to image the insect’s brain while in free flight, Erickson and her fellow researchers trick the creature. They keep the fly’s head immobile while allowing its body and wings to move, and then place it within a 360-degree cylindrical panel of LEDs that simulate what the fly would see while flying. Then researchers can look through the top of the insect’s head to see which neurons fire depending on how and in which direction a fly flies.

What Happens When We Go Under?

Caltech graduate student Jonathan D. Kenny studies anesthesia, a medical procedure that, despite its ubiquity and necessity, is shrouded in mystery. In the lab of Thanos Siapas, a professor of computation and neural systems, Kenny pursues the neural circuit dynamics of general anesthesia using very high-density electrodes implanted in the brains of animals. In this way, he can study the brain patterns of animals to understand what the brain is doing under anesthesia and to identify via brain waves the stages of recovery to full consciousness. After all, he says, people regain full awareness in stages rather than simply waking up and feeling normal again. “How can you tell if someone is conscious under anesthesia? Right now, the paradigm in the operating room is to look at things like your heart rate and blood pressure, but we know that anesthetics reduce consciousness by exerting their effects on your brain,” he says. “So, the notion is maybe we should observe brain activity.”

Like many Chen researchers, Kenny relies on the ascendant power of AI. Neural networks help Caltech researchers classify the noisy data from brain recordings to determine which brain wave states an animal or person is cycling through, with the goal of forming a complete picture of the brain under general anesthesia. Such a neural network may also be able to surface new brain states or combinations of brain states that a human researcher could not see amid all the noise.

AI Pushes Neuroscience Forward

Caltech’s Kortschak Scholars program, started with an endowment from businessman and Caltech trustee Walter Kortschak (MS ’82), supports computing mathematical science graduate students who pursue bold research ideas. For Jennifer Sun, that means the place where machine learning intersects with neuroscience.

Sun collaborates with the lab of Seymour Benzer Professor of Biology David Anderson, the Tianqiao and Chrissy Chen Institute for Neuroscience Leadership Chair and director of the Chen Institute, as researchers there investigate the neural control of social behaviors in mice by imaging in the brains of animals during periods of social interaction.

Identifying and scoring social behaviors from videos of the experiments had required many hours of painstaking manual annotation. Sun is training machine learning models to recognize mouse behavior automatically, allowing researchers to collect and process much larger volumes of behavioral data than was previously possible. “We want to help them reduce annotation effort and accelerate behavior experiments,” she says.

A Neural Network in a Dish

Suppose you grow neurons on a dish and then these brain cells begin to form connections to one another. Could they be trained to perform a computational task in much the same way researchers train machine learning algorithms? In Assistant Professor of Computational Biology Matt Thomson’s lab, graduate student Guruprasad Raghavan tries to do just that in a research project to fabricate “cortical computers.” This is a joint project with graduate student Varun Wadia from the lab of Doris Tsao, Professor of Biology, Chen Center for Systems Neuroscience Leadership Chair, Investigator at the Howard Hughes Medical Institute, and director of the Chen Center for Systems Neuroscience.

Just like training a dog to sniff bombs or training an AI to solve problems, building these cortical computers out of neurons requires positive and negative feedback to effectively tell the cells whether or not they are completing the task. “The biggest challenge,” Raghavan says, “is figuring out what is a reward or what is a punishment for neurons in a dish.” The team plans to find out whether feeding neurons dopamine as a reward is effective.

What the Brain and AI Can Teach Each Other

Artificial intelligence, for all its power and potential, struggles with problems that biological intelligence easily cracks. Sanghyun Yi, a Chen Graduate Fellow in the lab of Caltech psychology professor John P. O’Doherty, says that studying how the human brain solves what appear to be simple problems will illuminate better ways to design machine learning algorithms.

Consider a task like riding the subway. The task consists of multiple sub-tasks that should be solved in a very specific order: first buy a ticket to pass through the gates and board the train before it leaves. Human brains use memory and context to solve such a task in an efficient way. However, for an AI, it may not realize that the steps need to happen in a very specific order unless it runs through a multitude of simulations and figures out the problem by trial and error.

However, Yi says, the converse is also true: sometimes AI can reveal new insights into the inner workings of the brain. In a recently published study, Yi and colleagues sought to understand how the brain evaluates a piece of visual art. Starting with a computer vision analysis of the aesthetic elements and considering subjective human ratings about certain attributes in a work of art, the model could explain how brains would subjectively rate whether they like or dislike a painting or drawing.

Caltech’s David Prober is interested in how the brain regulates sleep. With zebrafish as his model, he has made several breakthrough discoveries about this still-mysterious state. With access to new research facilities and technology, his lab’s move into the new Chen Neuroscience Research Building could open the door to many more.

It is the one bodily need that cannot be denied. More than food or water or sex, sleep is the one activity that seemingly few animals can abstain from for very long.

“The simplest creature that’s been shown to sleep is the upside down jellyfish, and they don’t even have a brain,” says Caltech biology professor David Prober. “So, sleep must be doing something that’s really ancient and central and important. But we don’t know what that is.”

Improved knowledge about this mysterious state that occupies about a third of people’s lives could not only help resolve a central biological mystery, it could also help treat insomnia and other sleep disorders.

Prober is part of a growing cadre of researchers who are taking a fresh approach to investigating sleep by studying it in an…



Read More: Caltech: The Caltech Effect February 2021 – Neural Networking