A Learning-Based Approach to Artificial Sensory Feedback

Abstract

Proprioception plays an essential role in natural motor control, and we argue that it will serve an equally important function in artificial control of motor prosthetic devices. An artificial sensory feedback signal that could substitute for proprioception in a Brain-Computer Interface (BCI) must be sufficiently informative to be used alone when vision is not available (sensory substitution), and it should integrate with vision to improve motor performance when it is (sen-sory augmentation). Achieving these qualities with an artificial signal requires a high-bandwidth channel, which can be achieved with an invasive neural interface. With invasive electrode arrays, we can manipulate the activity of populations of neurons using intracortical electrical microstimulation (ICMS), effectively transmitting useful information directly to the neural circuits where it is needed. To date, the dominant strategy for encoding artificial somatosensation has been bio-mimetic-trying to replicate, at the single neuron level, the neural activity seen during natural sensory processing. Here, we argue for a different, though complementary , learning-based approach. We propose taking advantage of the natural plasticity of the sensorimotor system, and asking the brain to learn, de novo, an artificial input. We hypothesize that the statistical dependencies, such as temporal correlations, that will be imposed on a natural (vision) and an artificial sensory input (ICMS) will be enough to drive learning and, ultimately, integration of the two inputs. Therefore we suggest that such a learning-based approach can achieve sensory substitution and augmentation of vision, the two desired properties of an artificial sensory feedback signal for clinical motor neural prostheses.