Compact neural representations in co-adaptive Brain-Computer Interfaces

Brain-computer interfaces (BCIs) offer a unique method to study learning dynamics by defining a causal mapping between neural activity and movement. Examples include studying the acquisition of an arbitrary sensorimotor mapping or perturbing a learned mapping to study adaptation. Alternatively, the sensorimotor mapping (“decoder”) can be purposefully adapted in closed loop, as the subject controls the BCI, to improve performance. Beyond improving performance for applications, closed-loop decoder adaptation (CLDA) also allows the study of learning mechanisms under changing motor demands. Here, we studied neural plasticity in an adaptive BCI. Monkeys controlled a cursor towards targets using multi-unit activity from motor cortex. CLDA was used to both improve performance of initial decoders and to replace non-stationary units to maintain performance across days. We first analyzed the evolution of target encoding in both readout units and non-readout units (units recorded but not used for BCI control) using multiclass logistic regressions. During training, the readouts became significantly more predictive than the non-readouts, showing that credit was preferentially assigned to the readouts. Also, only a subset of units became necessary for efficient target encoding, which we define as a compact representation. This has not been reported with fixed decoders. To explore whether CLDA contributed to forming compact representations, we developed a novel neural network model with biologically plausible learning to simulate BCI tasks with and without decoder adaptation. Without CLDA, the representation was on average less compact, as determined by the impact of each readout neuron on task performance. Our findings reveal a potentially fragile neural representation dominated by only a few units that directly results from adapting sensorimotor mappings during learning. As a testbed that captures properties of experimental data, our model can be used to design new CLDA strategies to maintain performance while also creating redundant neural encodings needed for robust performance.