When controlling a motor prosthetic device or computer cursor using a brain-machine interface (BMI), users guide their movements relying solely on visual feedback. In contrast, our natural ability to plan and execute movements relies on feedback from multiple sensory signals, and in particular on proprioception---the sense of the body's position in space. Proprioception guides movement without the need for continuous visual monitoring and is integrated with vision, when both are available, to improve behavioral performance. Such multisensory integration appears to be learned from multisensory experience, a process that theoretically can be driven solely by the shared statistical structure of the inputs. To achieve naturalistic control over prosthetic devices, BMIs will likewise need a proprioceptive feedback signal, one that can ultimately be integrated with natural vision. Here, we demonstrate a novel, learning-based approach to artificial sensory feedback. We show that monkeys can learn to use, in a naturalistic way, a continuous, multi-dimensional, multi-channel intracortical microstimulation (ICMS) signal that encodes task-relevant feedback. After training with correlated ICMS and visual feedback, monkeys could perform a goal-directed reaching task without vision, using only the ICMS signal. Additionally, the animals integrated the natural and artificial sensory inputs, combining both into a minimum-variance sensory estimate of hand position relative to the target. The ICMS signal further resembled natural sensation in its plasticity.For example, when a misalignment was imposed between vision and ICMS, the ICMS estimate adapted back towards the rewarded visual cue. Furthermore, the monkey's valuation of the accuracy of the ICMS estimate was updated in response to trial conditions. When the ICMS is randomly, but not consistently, perturbed, the monkey's estimate using artificial feedback grows less precise. Finally, we discussed theoretical ways to improve the ICMS signal reliability by changing the parameters of the signal. Together these results demonstrate that a learning-based approach can be used to provide a rich artificial sensory signal, suggesting a new strategy for restoring proprioception to patients with BMIs as well as a powerful new tool for studying the adaptive mechanisms of sensory integration.

The sensory signals that drive movement planning arrive in a variety of `reference frames', and integrating or comparing them requires sensory transformations. I set out to examine how the different forms of sensory signals, and the transformations needed to compare them, affect the representation and integration of multimodal sensory information. I used a combination of human psychophysics and electrophysiological recordings from rhesus macaques to examine how visual and proprioceptive information are used for reach planning and execution.

The human experiment was designed to exploit stereotyped patterns of gaze- dependent reach errors to determine whether the reference frame representations for reach planning depend on the visual and proprioceptive sensory information available. The results of this experiment were interpreted with a model of reach planning in which the statistical properties of sensory signals and their transformations determine how these signals are used. I found that no single reference frame representation was adequate to explain the observed error patterns when visual and proprioceptive information were varied. Only by integrating movement plans across multiple reference frame representations was the model able to capture the observed error patterns (Chapter 1).

Taking the results from this model, I next looked for evidence of these multiple reference frame representations across different sensory modalities in sensory-motor cortical areas of the rhesus macaque (Area 5 and MIP). I found that neurons in these areas use mixed reference frame representations, which are consistent across reaches to targets specified by different sensory modalities (visual and/or proprioceptive targets, Chapter 2). Additionally, I found that integration of multimodal sensory signals in Area 5 and MIP emerges primarily across the population response rather than within individual cells' responses (Chapter 3). These findings are consistent with the model results showing that sensory information is integrated in multiple reference frame representations, regardless of the reference frame in which sensory information enters the nervous system. These results illustrate one way that the brain can represent and integrate sensory information arriving from different sensory modalities.

All ethologically-relevant behavior is dynamic, and biological in origin. To understand the neural underpinnings of motor behavior is then to quantify the limitations placed on organisms as they move through and interact with their environment. This thesis stands as an attempt to understand the way the brain processes information to prepare for movement, and how a network of neurons might give rise to that movement. I chose to investigate oculomotor tracking in non-human primates as a model behavior that allowed excellent spatial and temporal control over the stimulus. First, by carefully controlling the motion of a single target, and measuring the eyes' response to particular perturbations, I mapped out effective constraints on the way the brain processes target motion to guide smooth eye movements. The next set of experiments was motivated by the observation that no two tracking movements are alike, even when faced with identical stimuli. I establish the dynamic relationship between fluctuations in single neurons recorded in the frontal eye fields and concomitant behavioral fluctuations. Utilizing measurements of the covariation between simultaneously recorded neurons, I describe the properties of the network of cells which give rise said fluctuations, which is likely dominated by fluctuations in a small subset of the active population at any moment in time. Taken together, the work I've done first establishes that proper oculomotor tracking likely takes the form of a re-allocation of cortical resources, but one that may be dominated by the activity of a small number of cortical neurons.

The ability to make accurate movements in response to environmental stimuli is integral to the survival and well-being of many animal species. To that end, many of the movements we make on a daily basis have been practiced to the point of near-automatic accuracy. Even with intensive practice, however, repeated attempts at the same movement goal are inevitably variable at a fine scale. Traditional theories of motor control have attributed trial-by-trial variability in motor output to the failure of peripheral effectors, such as muscles and joints, to perfectly execute the motor command specified by the brain. More recent evidence suggests that at least some variability in motor output is correlated with variable neural responses in certain areas of the brain. This has led to the emerging view that trial-by-trial neural variability, rather than representing "noise" may be inherent to the "signal" that the brain uses to control motor output.

In this thesis, we used visually guided reaching as a model to study the relationship of neural variability in the dorsal premotor cortex (PMd) to movement variability. As described in Chapter 1, we found that for some PMd neurons, variability in planning-related activity, preceding movement onset, is correlated with trial-by-trial fluctuations in reach direction. We exploited these correlations to distinguish target- and movement-related neurons; our results suggest that PMd activity preceding movement is related to both the movement goal and the intended action. In Chapter 2, we quantified the fraction of movement direction variability predicted by planning-related activity of individual PMd neurons as well as of ensembles of 19-115 simultaneously recorded neurons. We found that preparatory activity of individual neurons typically predicted less than 1% of movement direction variability, while ensemble activity could predict up to 30-40%. Our results suggest that the movement planning process may be variable even under nominally identical task conditions, and that variability in the motor plan could give rise to measurable changes in movement direction across repeated trials.