State of the Science: Combining Neuroscience and AI

Enter a world of possibilities: generate ideas, get answers, have conversations, and create endless images. In many cases, artificial intelligence and large language models such as ChatGPT, Gemini, and Claude can appear almost human-like. However, this technology is still far from accurately representing the human brain. No matter how sophisticated the tools may seem, AI does not think for itself and its connections are no more complex than the brain. However, AI is in an important feedback loop with neuroscience. Advancing the questions being asked, accelerating technology, and accelerating discovery.

Building brain tools

Our senses, emotions, and memories make us unique human beings. How we take in information from the outside world and process it through our eyes, ears, nose, mouth, and fingertips remains many mysteries that neuroscientists and other researchers are trying to solve. And as we learn more, better model systems will be built, allowing us to answer new questions at an accelerated rate.

The discovery of how neurons in the auditory cortex respond to and process sounds to build better computational models of speech is consistent with studies of individual neurons being conducted by Dr. Samuel Norman Heigner, an associate professor of neuroscience and biostatistics at the University of Rochester. Norman-Haignere’s research leverages AI techniques to build better computational models that can predict how the human brain encodes complex sounds such as speech and music. His lab collects accurate, large-scale data from the human brain to train better computational models, while testing whether existing model calculations match the brain while also experimenting with models to generate predictions for the next generation of auditory neuroscience experiments.

A woman facing a microscope in a laboratory — Dr. Gabriella Stern, assistant professor of biomedical genetics and neuroscience in the medical center’s laboratory microscope.

The scientific loop between computational and model systems is exemplified by the FlyWire Connectome. This is a map of all the neurons and synaptic connections in the central brain of the fruit fly, the Drosophila melanogaster. Dr. Gabriela Stern, assistant professor of biomedical genetics and neuroscience, contributed to this research and has since used neural connections mapped to the connectome to accelerate research results in living Drosophila brains. For example, she was part of a team that discovered that there was overlap in different preference circuits, and it would have taken her years, if not her entire career, to discover whether a connectome-based computational model had not pointed researchers in the right direction. “We couldn’t complete the connectome without machine learning,” Stern said. “These findings showed that connectome-based models can predict non-intuitive circuit features and confirm them experimentally.”

gears of collaboration

From left: Karol Shimla, Dr. Krishnan Padmanabhan, Anna Kolstad.

Dr. Krishnan Padmanabhan, associate professor of neuroscience, conducts basic research on the relationship between the olfactory system and memory. This has applications in understanding several brain diseases, as olfactory disruption is often a symptom of neurodegenerative diseases. He and Dr. Julian Meeks, associate professor of neuroscience, are using machine learning to scrutinize large-scale data analysis to better understand this relationship.

Padmanabhan is also collaborating with Dr. Gourav Ghoshal, professor of physics and computer science, to develop new methods to explore what computation in the brain means and how it can provide insight into how the brain works. “Rather than studying specific regions of the brain, my group focuses on how patterns of activity in large-scale networks encode and convey meaningful information, and how these dynamics support coherent behavior,” Ghoshal said. “This has implications for both understanding the biological brain and developing new computational principles that may inform future AI systems. We’re seeing a lot of cross-pollination with AI in the broader field. Machine learning tools are increasingly shaping how neural data is analyzed, while ideas from neuroscience such as recurrent dynamics, modular composition, and energy-based structures continue to influence AI architectures.”

Snyder (left) and Hefner sit at computers in their research lab at the University of Rochester. — Mr. Snyder (left) and Mr. Hefner. // Photo courtesy of JA Fenster / University of Rochester

That deep understanding can be modeled with tools outside the brain, and recognizing patterns based on previously shared information is a strength of current AI systems. However, while AI can recognize patterns, it doesn’t have the ability to know where those patterns come from. Dr. Adam Snyder, assistant professor of neuroscience and brain and cognitive science, studies the relationships between cells in the biological brain. He and Dr. Ralph Hefner, associate professor of brain and cognitive sciences, study how learning progresses in the brain and how understanding is shared among many neurons. “This suggests that future systems may benefit from moving beyond just recognition to internal models that can explain, predict and infer inputs,” Snyder said.

Today’s AI is far from this point. But, as Hefner points out, their study is a step toward changing that. “No matter how advanced experimental techniques become, it will take a long time before we have enough data to constrain all the parameters needed for such a model. However, our study demonstrates the critical role of feedback connections that are largely missing from modern AI systems, including models of the brain.”

Machine learning yesterday and today

The roots of machine learning date back to the 1940s, when Warren McCulloch and Walter Pitts introduced the first mathematical models of neural networks. It has changed over the past 80 years and rapidly accelerated into a viable tool over the past several decades.
Progressing science.

In recent history, researchers known as the “godfathers” of AI, Jeff Hinton, Yoshua Bengio, and Yann LeCun, drew much inspiration from neuroscience to create their AI models. “Many of the biggest advances in AI have been inspired by the brain. The biggest challenge, in my opinion, is that many of today’s AI researchers are not interested in neuroscience and think AI has nothing to offer. But I strongly disagree,” said Christopher Kannan, Ph.D., associate professor of computer science and associate director of AI strategy at the Gergen Institute for Data Science and Artificial Intelligence.

Lean＞

Keinan’s research aims to mimic the brains of humans and other animals, allowing AI models to learn over time. “I take a lot of inspiration from the memory consolidation mechanisms that occur during sleep, especially the role of the hippocampus during non-REM sleep.
Effects of sleep (eye movement) and REM (rapid eye movement) on improving neural representation. ” In the visual system, Keinan’s research incorporates mechanisms associated with the prefrontal cortex into deep neural networks. He takes inspiration from neural architectures to overcome the limitations of today’s large-scale languages and vision language models, such as ChatGPT.

In 2025, the university joined the Empire AI Consortium. The Empire AI Consortium is a group of other public and private research institutions in New York State that aims to bridge the gap between researchers, public interest organizations, and small businesses to accelerate the development of artificial intelligence centered on the public interest.

Science accelerated by AI

In 2012, Miken Niedergaard, MD, DMSc, co-director of the university’s Center for Translational Neurology, discovered the glymphatic system. The discovery of this system, which removes waste from the brain during sleep, has accelerated our understanding of the relationship between body fluid flow, sleep, and human health and disease. It has also created a need in the neuroscience community to dig deeper into how neuroscience and other body fluids move through the brain.

In this way, mechanical engineering professor Dr. Douglas H. Kelley has become integrated into the research of neuroscience, a field of science that was unrelated to his work a decade ago. “Eurochester has the lowest boundaries between departments of any university I’ve ever seen,” Kelly says. “If I had her, I wouldn’t be studying brain fluid dynamics.” [Needergard] It wasn’t in Urochester. ” Kelly and his colleagues are now building machine learning models of brain fluid flow to measure the unmeasurable. “We simultaneously train models based on in-vivo measurements and the physics of fluid dynamics. “Our models can account for quantities such as pressure and estimate critical quantities such as flow much more accurately than other methods. AI will soon become more like simulation, a class of entire problem-solving approaches rather than a single scientific tool,” Kelly said. New methods are being invented all the time, each tailored to specific problems. One way neuroscience and fluid dynamics are impacting AI is by pinpointing important issues where new AI methods are needed and can make a real difference. ”

Translate research into clinical practice

Advances in machine learning for analyzing large data sets are accelerating research by Frank Garcea, Ph.D., assistant professor of neurosurgery, and Michelle Janelsins, Ph.D., MPH, Joan and Gary Morrow Endowed Distinguished Professor of Cancer Supportive Care at Wilmot Cancer Institute. Working with cancer patients through a translational brain mapping program, Garcea, Janelsins, and M.D./PhD student Emma Stroderman found that brain networks in the right hemisphere rewire in response to tumors in the left hemisphere, and that patterns of rewiring before surgery can predict whether a patient will be unable to speak fluently after surgery.

From left: Stroderman, Garcea, Janel Shins

Garcea believes the technology is a promising clinical tool, but stresses the need for caution. “Changing the way we use this technology to predict outcomes creates real risks,” he said. “For example, not all rewiring is equally important. How the visual networks in the right hemisphere are rewired before surgery does not predict whether a patient will be unable to speak fluently after surgery, but it does predict how the language networks in the right hemisphere will be rewired.”

He emphasizes that each advancement in technology creates new risks to consider in his work and in science as a whole. “Careful evaluation of training data and predictive accuracy is essential, highlighting the importance of keeping human expertise at the center of the use of these tools in both research and clinical decision-making.”

Source link