M8 Brain-machine interfaces

Topic

A brain-machine interface records the neuronal activity of a population of neurons and treats that activity—as described by the population encoding and decoding model—as a collective code representing the person's motor or cognitive intention at any given moment. Decoding this recorded neuronal activity entails inverting the population code to determine, based on the simultaneous activity of many neurons, the specific intention the person is attempting to express.

An adaptive linear decoder performs this decoding task by combining the activity of the neuronal population using a set of specific weights—one for each recorded neuron—to produce a decoded signal that estimates the person's intention at any moment. Initially, these weights do not combine the neuronal activity effectively, and the resulting decoded signal deviates from the actual target; however, a supervised learning rule adjusts the weights session by session, moving in a direction that reduces the error between the decoded signal and the known target intention established during the decoder's training. The more this adjustment is repeated, the smaller the error becomes, and the more faithfully the decoded signal reflects the person's actual intention.

This decoded signal goes beyond mere estimation: it directly controls an external effector, such as a cursor or a robotic arm. The movement of this external effector, in turn, generates a sensory consequence that is returned to the person via artificial sensory feedback, allowing them to perceive the result of their decoded intention and make immediate corrections if the outcome was not the one intended. This entire sequence—from recorded neuronal activity to the decoded signal, from the decoded signal to the external effector, and from the external effector back to the person via artificial sensory feedback—closes the control loop of this brain-machine interface. Repeated use of this brain-machine interface not only adjusts the decoder's weights but also induces—as described by the plasticity model—cortical reorganization within the user, mirroring the process outlined in the M05 model. As the user repeatedly employs the interface, the cortex learns on its own to modulate the activity of that specific neuronal population to make it easier to decode; thus, the decoder adapts to the cortex through supervised weight learning while the cortex adapts to the decoder through its own cortical reorganization. Together, these dual adaptations result in superior device control compared to what either process could achieve in isolation.

The performance of this brain-machine interface is quantified in two ways. The first is the information transfer rate, calculated using the method introduced by Wolpaw—based on the same information theory underlying the population encoding and decoding model. This metric encapsulates, in a single figure, how many distinct options the interface can distinguish, the accuracy of its selections, and the number of decisions made per minute. The second measure is decoding fidelity, which compares the decoded signal against the intended action—much like the supervised learning rule used during training. This fidelity typically degrades over the course of a prolonged session, as user fatigue makes it difficult to maintain the clear, consistent neuronal activity the interface requires to accurately decode the user's intent.

ID:4280