Thalamocortical credit routing from the basal ganglia and cerebellum: anatomical constraints and circuit principles

To understand how an animal learns to choose the right action for a reward, scientists have long focused on two deep brain structures: the basal ganglia and the cerebellum. These regions act as learning engines. The basal ganglia help decide which action is worth pursuing based on past rewards, while the cerebellum fine-tunes the timing and precision of movements. Both structures receive instructions from the cerebral cortex, the brain's outer layer responsible for complex thought and planning. After processing this information, they send signals back to the cortex to update future behavior. This creates a continuous loop where the brain constantly checks its own performance. However, a critical piece of this puzzle has remained unclear: how do these learning signals travel back to the specific groups of neurons in the cortex that actually generated the action? Without a precise delivery system, the brain would not know which neurons to strengthen or weaken, making learning inefficient or impossible.

A team of researchers in Japan has now mapped this delivery route in mice, revealing a sophisticated system that ensures learning signals reach the exact right destinations. They discovered that while the basal ganglia and cerebellum send their messages through largely separate channels within the thalamus—a central relay station in the brain—these channels merge again when they reach the cortex. Specifically, the researchers found that signals from both systems converge on the very top layer of the secondary motor cortex, a region crucial for planning and learning new skills. This convergence allows the brain to combine different types of learning information and direct them to the specific neurons responsible for an action. By combining detailed brain imaging with computer simulations, the team showed that this anatomical arrangement is essential for the brain to assign "credit" correctly, ensuring that the neurons that contributed to a successful action are the ones that get reinforced.

The researchers began by tracing the physical pathways in the brains of mice to see where these learning signals go. They injected tiny amounts of fluorescent tracers into the basal ganglia and the cerebellum and watched how the signals traveled through the thalamus to the cortex. They found that the neurons in the thalamus receiving input from the basal ganglia were mostly located in one area, while those receiving input from the cerebellum were in a different area. This confirmed that the two systems remain largely separate as they pass through the thalamus. However, when the researchers looked at where the axons—the long fibers that carry signals from the thalamus to the cortex—ended up, the picture changed. Despite starting from different places in the thalamus, the axons from both systems shared a striking destination: the uppermost layer of the secondary motor cortex. This layer contains the far-reaching tips of the nerve cells that control movement. The researchers also found a small area in the thalamus where signals from both systems mixed, but the most significant finding was this shared arrival point in the cortex.

To understand why this shared destination matters, the team built a detailed computer model of the entire brain circuit, including the cortex, thalamus, basal ganglia, and cerebellum. In their simulation, they tested how the brain learns when signals are routed correctly versus when they are scrambled. They found that learning worked best when the feedback from the basal ganglia and cerebellum was sent back to the specific groups of neurons that had helped generate the action. If the signals were sent to the wrong neurons, learning failed. The model also revealed that the brain uses two different ways to send these signals back. One type of signal acts like a volume knob, making the existing activity of a neuron stronger or weaker. The other type acts like a direct push, adding new activity to the neuron. The researchers discovered that the brain needs both of these mechanisms working together to learn efficiently. The "volume knob" effect, which happens in that upper layer of the cortex, helps the brain amplify the right signals, while the direct push helps drive the action forward.

The study also explored how the brain figures out which neurons should receive these learning signals in the first place. In the computer model, the researchers showed that the brain can learn to align its feedback by observing its own activity during periods of rest. By comparing how active a neuron was with how likely it was to trigger an action, the brain can adjust its connections so that learning signals are sent to the right place. This process, which they called local calibration, allowed the model to recover its learning ability even if the initial connections were not set up perfectly. Furthermore, the researchers tested whether this system could handle complex tasks where multiple neurons needed to be trained at the same time, such as in brain-machine interface experiments where a person learns to control a computer cursor with their thoughts. They found that when the brain sends a single, global reward signal to the whole network, it struggles to train multiple targets simultaneously if those targets need to move in opposite directions. However, when the feedback is routed through separate, parallel channels, the brain can successfully train multiple neurons at once. This suggests that the brain's ability to learn complex, multi-part tasks depends on having these distinct, parallel pathways that can deliver credit to specific groups of neurons without interference.

These findings offer a new way to view how the brain learns. Rather than seeing the basal ganglia and cerebellum as independent systems that simply add their outputs together, the researchers propose that they work as a coordinated team that uses a shared highway to deliver learning signals. This highway, which converges on the top layer of the motor cortex, allows the brain to integrate different types of information and direct it precisely where it is needed. The study suggests that the physical layout of these connections is not just a passive wiring diagram but an active component of the learning process itself. By ensuring that learning signals reach the specific neurons responsible for an action, the brain can efficiently refine its behavior and adapt to new challenges. This work provides a concrete anatomical and computational framework for understanding how the brain solves the problem of credit assignment, a fundamental challenge in learning that applies not only to animals but also to the development of artificial intelligence systems. The researchers' results, derived from both direct observation in mice and rigorous computer simulations, highlight the importance of laminar convergence and modular feedback in enabling the brain to learn complex behaviors.