MIT researchers have discovered that the brain utilizes precise, individualized teaching signals to guide learning, a mechanism strikingly similar to the backpropagation methods used in artificial intelligence. The findings, published on February 25 in the journal Nature by a team led by Mark Harnett at the McGovern Institute for Brain Research, demonstrate that individual neurons receive specific error signals to adjust their activity, rather than relying on broad, inefficient chemical broadcasts. Traditional reinforcement learning in the brain often involves neuromodulators like dopamine, which release general signals affecting large groups of neurons simultaneously. While effective for general adaptation, this approach lacks the specificity needed for efficient, targeted learning. In contrast, machine learning systems use backpropagation to compute precise error signals for each connection within a network, allowing for rapid optimization. For decades, neuroscientists have debated whether biological brains could achieve similar precision, given the physical constraints of living tissue compared to software algorithms. The primary obstacle has been the difficulty in identifying vectorized instructive signals that dictate exactly which neurons should increase or decrease activity. To overcome this challenge, Harnett and first author Valerio Francioni developed a novel brain-computer interface (BCI) experiment involving mice. They selected eight to ten specific neurons in the mouse cortex and linked their activity to a visual interface. The task required some neurons to increase their firing rates while others needed to decrease activity to succeed. Success was rewarded with a sugary treat, providing clear sensory feedback on performance. By monitoring the dendrites of these specific neurons daily using high-resolution microscopy, the team tracked how activity changed in response to rewards or failures. They observed that neurons tasked with increasing activity received instructive signals signaling them to ramp up, while those required to decrease activity received opposing signals. Crucially, when the researchers experimentally blocked these dendritic signals, the mice failed to learn the task, confirming that these vectorized signals are essential for the learning process. This provides the first direct biological evidence that the cortex employs neuron-specific error signals similar to those in deep learning networks. This discovery bridges the gap between neuroscience and machine learning, offering a powerful new framework to study how the brain learns. The ability to translate concepts from artificial intelligence into biological experiments allows researchers to test hypotheses about cortical learning with unprecedented accuracy. Harnett and his team plan to use this approach to explore how different brain regions learn and to develop more sophisticated, brain-inspired AI models. As Francioni noted, the capacity to directly translate mathematical models of learning into the biology of the brain marks a significant new beginning for understanding neural function and creating advanced neural technologies. The study not only clarifies the mechanics of individualized learning in the brain but also encourages closer collaboration between neuroscientists and AI developers to refine both biological understanding and computational models.