Bella Chin also contributed to this story.
When a dog learns a new trick, something amazing happens within their brain: The neurons literally rewire themselves. And as the dog begins to nail this trick time and time again, its brain goes through an optimization process in which various neuronal parameters, like synaptic weight (the strength of the connection between two neurons), are tweaked to minimize performance error.
Rutgers University researchers used National Science Foundation (NSF) ACCESS allocations on Bridges-2 at the Pittsburgh Supercomputing Center, as well as the Texas Advanced Computing Center (TACC) allocated system, Frontera, to build an artificial neural network that follows several real brain rules. Their goal was to make a model that’s easier to compare with actual biology, so it can help explain how learning is coordinated across different layers of the brain.
The team built the network with four biological constraints in mind: different cell types, excitatory and inhibitory signaling, continuous signaling through dendrites, and non-symmetric connections between neurons. In plain terms, that means the model doesn’t treat every neuron as identical, nor does it rely on the simplified wiring assumptions common in standard AI models.
“One major change was creating separate excitatory and inhibitory neurons in each layer of the network,” said Aaron Milstein, an assistant professor of neuroscience and cell biology at Rutgers University. “That helped the Frontera- and Bridges-2-generated model show brain-like behavior, including competition between neurons, sparse activity and stronger selectivity in some cells.”
The researchers also added dendrite-based processing, since real neurons receive and filter signals in more than one place at once. They then introduced a learning method called dendritic target propagation, along with a local learning rule that removed the need for perfectly symmetrical connections.
The result is a biologically constrained model – thanks to NSF ACCESS allocations on Bridges-2 and the use of TACC Frontera – that still learns effectively, while staying closer to how neurons actually work. This makes the system useful for testing ideas about neuroplasticity and for making predictions that future experiments can check.
Aaron Milstein, Rutgers University
This research was published in Cell Reports. Authors include Milstein, Alessandro R. Galloni, Ajay Peddada and Yash Chennawar.
If you’re a researcher in need of compute power for your simulations, you can get started with ACCESS here.
Resource Provider Institution(s): Pittsburgh Supercomputing Center (PSC)
Resources Used: Bridges-2
Affiliations: University of Illinois Urbana-Champaign
Funding Agency: NSF
Grant or Allocation Number(s): BIO250014
The science story featured here was enabled by the U.S. National Science Foundation’s ACCESS program, which is supported by National Science Foundation grants #2138259, #2138286, #2138307, #2137603, and #2138296.