Available online at www.sciencedirect.com ScienceDirect Brain–computer interfaces for dissecting cognitive processes underlying sensorimotor control 1,3 2,3 3,4,5 Matthew D Golub , Steven M Chase , Aaron P Batista 1,2,3 and Byron M Yu Abstract formidable challenge for neuroscientists seeking a more complete understanding of the neural mechanisms un- Sensorimotor control engages cognitive processes such as derlying sensorimotor control. prediction, learning, and multisensory integration. Understanding the neural mechanisms underlying these A common paradigm for studying sensorimotor control is cognitive processes with arm reaching is challenging because arm reaching (Figure 1, left). Even the simplest of arm we currently record only a fraction of the relevant neurons, the movements emerge from a complex set of neural, muscular arm has nonlinear dynamics, and multiple modalities of sensory and skeletal systems. Movement generation involves mul- feedback contribute to control. A brain–computer interface tiple distinct cortical areas that project to the spinal cord, (BCI) is a well-defined sensorimotor loop with key simplifying numerous striatal and cerebellar loops, and several brain- advantages that address each of these challenges, while stem and thalamic nuclei [10]. However, in typical experi- engaging similar cognitive processes. As a result, BCI is ments, we can monitor only a tiny fraction of the hundreds becoming recognized as a powerful tool for basic scientific of thousands of neurons that project to motoneuron pools, studies of sensorimotor control. Here, we describe the benefits and it is often unknown whether the recorded neurons of BCI for basic scientific inquiries and review recent BCI project to the spinal cord or affect behavior only indirectly. studies that have uncovered new insights into the neural As a result, it is difficult to causally attribute behavioral mechanisms underlying sensorimotor control. Addresses changes to specific changes in the recorded neural activity. 1 Department of Electrical and Computer Engineering, Carnegie Mellon Furthermore, the arm is a multi-jointed structure actuated University, United States by dozens of muscles [11], and because of these complexi- 2 Department of Biomedical Engineering, Carnegie Mellon University, ties, the arm’s nonlinear dynamics are not typically mea- United States 3 sured in studies of arm reaching. In addition, sensory Center for the Neural Basis of Cognition, Carnegie Mellon University & University of Pittsburgh, United States feedback about the arm movement is carried through 4 Department of Bioengineering, University of Pittsburgh, United States multiple sensory modalities, including vision and proprio- 5 Systems Neuroscience Institute, University of Pittsburgh, United States ception, that have different latencies and need to be combined [12]. Although visual feedback about the arm can be readily manipulated [8,12–14], proprioceptive feed- back cannot be decoupled from movement as easily [15]. Current Opinion in Neurobiology 2016, 37:53–58 This review comes from a themed issue on Neurobiology of behavior How can we obtain a more complete understanding of the cognitive processes underlying sensorimotor control in Edited by Alla Karpova and Roozbeh Kiani light of this daunting complexity? Perhaps we could gain traction if we could simultaneously record neural activity from multiple brain areas in the motor system, including http://dx.doi.org/10.1016/j.conb.2015.12.005 all neurons that directly drive movement; if we could # 2015 Elsevier Ltd. All rights reserved identify and reversibly reprogram the precise mathemati- cal relationship between neural activity and movement; and if we could independently alter different modalities of sensory feedback. In this review, we describe how brain–computer interfaces Introduction (BCIs) provide a simplified, well-defined, and easily ma- Successful sensorimotor control requires the coordination nipulated experimental paradigm that facilitates the basic of multiple cognitive processes. On a moment-by-mo- scientific investigation of the cognitive processes engaged ment basis, the brain integrates various sources of sensory during sensorimotor control (Figure 1, right). A BCI creates information [1–3], selects and plans upcoming move- a direct mapping between recorded neural activity and the ments [4–6], internally predicts the consequences of movement of a device, such as a computer cursor (or robotic motor commands [7], and adapts to compensate for limb) [16,17 ,18–22,23 ] and substantially simplifies the changes in the body and environment [8,9]. Addressing complexities described above (Table 1). Although neurons the complexity of these interconnected processes poses a throughout the brain can indirectly influence movement, www.sciencedirect.com Current Opinion in Neurobiology 2016, 37:53–58 54 Neurobiology of behavior Figure 1 goal goal neural activity spinal cord neural activity BCI brain brain and arm mapping arm cursor movement movement vision vision arm proprioception proprioception movement Current Opinion in Neurobiology Conceptual illustration of the sensorimotor control loop for arm reaching (left) and BCI (right) movements. only the activities of experimenter-chosen output neurons the recorded neural activity because there are unre- directly drive BCI cursor movements. Thus, all aspects of corded neurons that can directly drive behavior. In behavior must be expressed in the activity of these contrast, for BCI, all neurons that directly drive behav- recorded output neurons, and it is possible to causally ior are recorded, by construction. Thus, it is possible to interpret the role of each neuron in behavior since the causally attribute changes in behavior to the activity of mapping between neural activity and cursor movement is specific neurons. completely known to and specified by the experimenter. Cursor dynamics can be defined by the experimenter to be This property of BCI is particularly well exemplified by linear, and they can be altered as desired. Furthermore, we single-unit operant conditioning studies. In operant con- can flexibly manipulate sensory feedback because propri- ditioning, the subject’s task is to learn to volitionally oceptive feedback is not hard-wired to cursor movement, modulate neural activity to specified levels using real- as it is for arm movement. As a result of these simplifica- time visual [29,30] or auditory [31 ] feedback. The visual tions, we and others have begun to leverage BCI as a or auditory feedback represents the BCI ‘behavior’. Op- powerful experimental paradigm for addressing basic sci- erant conditioning is particularly valuable for studying entific questions about sensorimotor control. Although (internal) cognitive processes because it allows the ex- similar concepts apply to EEG-based (e.g. [24]) and perimenter to manipulate neural activity that has, under ECoG-based (e.g. [25]) BCI, we focus on intracortical ordinary circumstances, only an indirect relationship to BCI in this review. Previous reviews have described use externally measurable variables. This approach was pio- of BCI for addressing scientific questions [26–28]. Here, we neered by Fetz [30] in the motor cortex, and later adopted focus on the key simplifications offered by BCI and de- in a large body of studies [32–37]. More recently, studies scribe the scientific insights that have emerged by leverag- have demonstrated the importance of operant condition- ing each simplification. ing for studying the neural substrates of cognitive pro- cesses, including spatial [31 ] and object-based [38] Monitoring all neurons that directly drive attention, that are involved in sensorimotor control. In movement particular, Schafer and Moore [31 ] found that volitional In arm reaching studies, it is typically not possible to changes in frontal eye field (FEF) activity are associated attribute every aspect of behavior to specific features of with selective visual attention. Table 1 Comparison of BCI control to arm reaching. Bold items indicate entries that make BCI a simplified, well-defined and easily-manipulated system for studying sensorimotor control Arm reaching BCI Effector Arm Cursor or robotic limb # of non-output neurons Millions Millions # of output neurons Thousands (only a subset are recorded) Tens-to-hundreds (all are recorded) Neuron-to-movement mapping Unknown Known Effector dynamics Difficult to measure, nonlinear Known, can be linear Sensory feedback Tied to arm Flexibly manipulable Current Opinion in Neurobiology 2016, 37:53–58 www.sciencedirect.com BCI for studying sensorimotor control Golub et al. 55 Distinguishing between output and non- enable selection of appropriate neural activity patterns output neurons and internal prediction to compensate for sensory feed- Beyond monitoring all neurons that directly drive behavior back delays [8]. We found that cursor movement errors in a BCI (output neurons), it is also possible to simulta- can be explained by a mismatch between the internal neously monitor additional neurons that are not explicitly model and the BCI mapping [50 ]. By using a linear BCI mapped to behavior (termed non-output neurons). This is mapping, we could focus on the family of linear internal advantageous for investigating whether and how the activ- models, which facilitated the identification of internal ity of output neurons of sensorimotor control differs from models from neural activity. Another convenient property that of non-output neurons. For example, it may be that the of a linear BCI