Image/ChatGPT
Biomedical engineers at USC Viterbi School of Engineering are embarking on a bold new project to unlock how the brain learns and processes information. The work, funded by the Defense Advanced Research Projects Agency (DARPA), is led by Dong Song, director of the USC Neural Modeling and Interface Lab and associate professor of biomedical engineering and of neurological surgery at Keck School of Medicine of USC. This new DARPA INSPIRE (Investigating how Neurological Systems Process Information in Reality) project will delve into the mysteries of Long-Term Synaptic Plasticity (LTSP), the fundamental mechanism by which our brains adapt and form memories.
The brain’s incredible ability to learn and remember is largely attributed to LTSP —the changes in the strength of connections between neurons that persist over time. However, the exact nature of LTSP and how it functions during real-world behaviors remains poorly understood. The primary challenge has been the absence of a method to directly investigate LTSP using the brain’s natural “operational signals” (spikes) in animals as they learn.
“It is a very difficult problem to solve,” Song said. “First, you need to be able to get the signal. You need a multielectrode array and an in vivo recording technique that allows you to record brain signals during learning and memory formation, which is often difficult.”
Existing studies often rely on indirect methods, artificial electrical stimulation, or “snapshot-like” comparisons that fail to capture the dynamic evolution of synaptic strength during active learning. This project aims to bridge that critical gap and create a computational tool to measure synaptic plasticity.
“If you think about our neurons, they receive many inputs to generate their outputs. A lot of things are going on and it’s very difficult for you to tease out which neurons are connecting to which neuron, how strong the synapse strings are, whether they change over time, and whether that change over time is governed by learning,” Song said.
A Novel Computational Approach
Postdoctoral researcher Xiang Zhang and associate professor of biomedical engineering Dong Song.
Song and his team (postdoctoral researcher Xiang Zhang and Ph.D. student Chen Sun) plan to develop and apply a sophisticated computational modeling approach to identify population-level LTSP directly from spiking activity in animals engaged in various behavioral tasks. This will be the first time such a direct investigation is possible.
“This strategy will allow us to go directly to the animals, as they are learning and forming memories, capture their working signals, and from that signal we’ll use this machine learning model to identify what the long-term synaptic plasticity is like,” Song said.
Looking Ahead: Biomarkers and Brain-Inspired AI
The 12-month project will enable foundational knowledge that could impact the future diagnosis and treatment of learning and memory disorders. For the first time, scientists will be able to directly investigate how the brain’s internal signals alter synaptic strength and reorganize neural networks for information processing and storage. The identified population-level synaptic learning rules could serve as new biomarkers for understanding both normal brain function and various neurological disorders.
“We will directly answer the question of whether long-term synaptic plasticity is the underlying mechanism of learning and memory formation,” Song said. “If that’s true, when an animal is actually learning, I can see it. If they are not learning, I shouldn’t see it. We can directly test that hypothesis. In addition, we can use it as a biomarker of brain signals to see which is a normal brain and which is a pathological brain.”
Song added that these biologically realistic learning rules developed by the project could provide critical insights for developing the next generation of artificial intelligence systems that more closely mimic the brain’s efficient learning capabilities. The work could also be instrumental for the hippocampal memory prostheses that Song and his collaborators have been developing for many years to ultimately restore lost memories in patients with neurological disorders.
“For the prostheses, developing this learning rule can render the whole system time-varying and adaptive. Our previous prostheses cannot change over time. We simply capture the stationary input-output functions from individual patients,” Song said. “But here, if we can identify the learning rule, we can potentially use it to make more intelligent prostheses that form new memories and learning new things just like the brain does.”
Published on July 31st, 2025
Last updated on July 31st, 2025