MIT Finds Neurons Get Personalized Teaching Signals
Researchers at the MIT McGovern Institute discovered that neurons receive highly specific "teaching signals" as we learn new skills, acting like personalized feedback to optimize how each neuron adapts during the learning process. The breakthrough has major implications for educational methods, cognitive training, and rehabilitation after injury. The discovery underscores the importance of individualized, responsive instruction in every learning environment.
- The study was led by Mark Harnett, a McGovern Institute investigator, with Valerio Francioni, a former postdoctoral researcher in Harnett's lab, as the paper's first author. Harnett's research focuses on how the biophysical features of individual neurons, particularly their dendrites, contribute to the brain's computational power. - To isolate and study the learning signals, researchers used a brain-computer interface (BCI) to train mice. The mice learned to control a visual stimulus on a screen by consciously modulating the activity of two specific, intermingled populations of four to five neurons each in their retrosplenial cortex. - This research provides the first biological evidence of a "vectorized instructive signal" in the cortex. This is a learning mechanism similar to the backpropagation algorithm used in AI, which sends targeted, specific error signals back to individual "neurons" in the network so they can adjust their connections. - This discovery helps solve the "credit assignment problem" in neuroscience—the question of how the brain knows which specific neurons or synapses to credit for a successful or failed outcome. The vectorized signals provide this precise, cell-by-cell feedback. - Previously, a dominant model for learning involved neuromodulators like dopamine, which broadcast a global reinforcement signal across large groups of neurons. This method is considered less efficient because it doesn't provide discriminating feedback to individual neurons based on their specific contribution to the task. - The experiment found these specific teaching signals in the dendrites, the tree-like structures on neurons that receive inputs. When the researchers used optogenetics to disrupt these dendritic signals, the mice failed to learn the BCI task, confirming the signals' instructive role. - This new understanding of a precise, backpropagation-like learning mechanism in the brain is expected to fuel further collaboration between neuroscience and machine learning research. It provides a new framework for investigating how different brain regions learn and for developing more brain-inspired AI.
