The hippocampus is one of the most important parts of our brain. It’s the seahorse-shaped structure in the center of the organ that is responsible for forming our new memories.
For the first time, USC Viterbi School of Engineering researchers will develop a highly-complex computational model of the hippocampus designed to function exactly like the real thing. The project has secured a grant of nearly $1.2 million from the NIH Brain Research Through Advancing Innovative Neurotechnologies (BRAIN) Initiative: Theories, Models and Methods for Analysis of Complex Data from the Brain.
Dong Song, a research associate professor in USC’s Department of Biomedical Engineering, will lead the hippocampus model project, along with BME post-doctoral researcher Gene Yu (now at Duke University), David Packard Chair in Engineering Ted Berger, and Dean’s Professor of Biomedical Engineering Vasilis Marmarelis. The research team will also include Michael Bienkowski, an assistant professor of Physiology and Neuroscience from Keck School of Medicine of USC.
Song was part of a team awarded a $6 million BRAIN Initiative grant in 2020 to develop polymer electrodes to monitor brain signals. He said that while his previous BRAIN initiative grant was all about collecting data from the brain, the latest project would draw on decades of collected brain data and knowledge to build a computational model that could mimic the function of a real hippocampus. The virtual model would then be cross-referenced with a real-life rodent hippocampus using machine learning to ensure its accuracy, in order to better understand the biology of the hippocampus and generate experimentally testable hypotheses about how it works.
“The main function of the hippocampus in higher-level mammals, such as humans, is episodic memory, especially the memory of your past experiences. It’s the autobiographical memory, as opposed to learning new skills,” Song said.
“In lower-level mammals, the main function of the hippocampus is spatial navigation, and this is the function I will be looking at in this proposal, using rats as the animal model,” he said.
Song said that while the hippocampus was one of the better-known areas of the brain in terms of its anatomy and function, there was still a great deal that the research community does not know about how its neurons work in concert to form higher-level cognitive function such as navigation and memory.
“To me that’s the Holy Grail of neuroscience–how those basic mechanisms work together to create a very complex function,” Song said.
Song and his collaborators will be using a machine learning framework known as a generative adversarial network, which until now has been used for image processing. The framework uses two models, which must work together to create a more accurate model.
One model is the “generative” realistic model of the existing world—in this case, the hippocampus—developed from neuroscience knowledge and hypotheses. The modeling framework would then look for discrepancies between this model and the real hippocampus using a “discriminative” machine learning model.
“On the one hand, we have rats navigating their environment and generating all sorts of signals about spatial representation. We’ll also have a model of a rat hippocampus navigating a virtual environment, just like the rat is doing, and generating signals. We want to know whether those two signals are sufficiently similar,” Song said.
“If there is a discrepancy, we will tune the realistic model to better fit the data. Then we do that again and again, iteratively, until the machine learning model can no longer distinguish between the model and the true biological system,” Song said.
Song said that such a computational modeling framework was sorely needed in neuroscience, based on his consultations with leading researchers in the field. “We need a general computational framework for integrating biological knowledge, hypotheses, and large-scale input-output data to gain deeper understanding of cognitive functions” Song Said. The project would also benefit from the expertise of Bienkowski, who would provide detailed anatomical data of the hippocampus’ structure, which would be crucial for forming the realistic model.
The application of the model, Song said, would be primarily to advance our understanding of the normal functions and processes of the hippocampus—how it processes information into signals from one brain region to another, and how spatial information is encoded in the activities of its neurons.
“But it can also provide a tool to understand neurological and cognitive disorders—so if you’re studying a memory disorder, you can use this platform to understand how that memory deficit is formed,” Song said.
For Song, combining biological modeling and machine learning to study the hippocampus has been a career-long passion. This approach was first attempted when he was a graduate student working under Marmarelis and Berger, where he modeled one individual synapse in the hippocampus.
“It was a tiny part of the neuron and it got very interesting results,” Song said. “I’m super excited that after almost 20 years I was able to expand this research from a tiny synapse to the whole hippocampus, which includes millions of neurons and billions of synapses.”