The Computational and Neural Bases of Context-Dependent Learning.

Flexible behavior requires the creation, updating, and expression of memories to depend on context. While the neural underpinnings of each of these processes have been intensively studied, recent advances in computational modeling revealed a key challenge in context-dependent learning that had been largely ignored previously: Under naturalistic conditions, context is typically uncertain, necessitating contextual inference. We review a theoretical approach to formalizing context-dependent learning in the face of contextual uncertainty and the core computations it requires. We show how this approach begins to organize a large body of disparate experimental observations, from multiple levels of brain organization (including circuits, systems, and behavior) and multiple brain regions (most prominently the prefrontal cortex, the hippocampus, and motor cortices), into a coherent framework. We argue that contextual inference may also be key to understanding continual learning in the brain. This theory-driven perspective places contextual inference as a core component of learning.

Contextual inference underlies the learning of sensorimotor repertoires.

ASBTRACT: Humans spend a lifetime learning, storing and refining a repertoire of motor memories. For example, through experience, we become proficient at manipulating a large range of objects with distinct dynamical properties. However, it is unknown what principle underlies how our continuous stream of sensorimotor experience is segmented into separate memories and how we adapt and use this growing repertoire. Here we develop a theory of motor learning based on the key principle that memory creation, updating and expression are all controlled by a single computation-contextual inference. Our theory reveals that adaptation can arise both by creating and updating memories (proper learning) and by changing how existing memories are differentially expressed (apparent learning). This insight enables us to account for key features of motor learning that had no unified explanation: spontaneous recovery1, savings2, anterograde interference3, how environmental consistency affects learning rate4,5 and the distinction between explicit and implicit learning6. Critically, our theory also predicts new phenomena-evoked recovery and context-dependent single-trial learning-which we confirm experimentally. These results suggest that contextual inference, rather than classical single-context mechanisms1,4,7-9, is the key principle underlying how a diverse set of experiences is reflected in our motor behaviour.

Fast Feedback Responses to Categorical Sensorimotor Errors That Do Not Indicate Error Magnitude Are Optimized Based on Short- and Long-Term Memory.

Skilled motor performance depends critically on rapid corrective responses that act to preserve the goal of the movement in the face of perturbations. Although it is well established that the gain of corrective responses elicited while reaching toward objects adapts to different contexts, little is known about the adaptability of corrective responses supporting the manipulation of objects after they are grasped. Here, we investigated the adaptability of the corrective response elicited when an object being lifted is heavier than expected and fails to lift off when predicted. This response involves a monotonic increase in vertical load force triggered, within ∼90 ms, by the absence of expected sensory feedback signaling lift off and terminated when actual lift off occurs. Critically, because the actual weight of the object cannot be directly sensed at the moment the object fails to lift off, any adaptation of the corrective response would have to be based on memory from previous lifts. We show that when humans, including men and women, repeatedly lift an object that on occasional catch trials increases from a baseline weight to a fixed heavier weight, they scale the gain of the response (i.e., the rate of force increase) to the heavier weight within two to three catch trials. We also show that the gain of the response scales, on the first catch trial, with the baseline weight of the object. Thus, the gain of the lifting response can be adapted by both short- and long-term experience. Finally, we demonstrate that this adaptation preserves the efficacy of the response across contexts.SIGNIFICANCE STATEMENT Here, we present the first investigation of the adaptability of the corrective lifting response elicited when an object is heavier than expected and fails to lift off when predicted. A striking feature of the response, which is driven by a sensory prediction error arising from the absence of expected sensory feedback, is that the magnitude of the error is unknown. That is, the motor system only receives a categorical error indicating that the object is heavier than expected but not its actual weight. Although the error magnitude is not known at the moment the response is elicited, we show that the response can be scaled to predictions of error magnitude based on both recent and long-term memories.

Interaction between decision-making and motor learning when selecting reach targets in the presence of bias and noise.

Motor errors can have both bias and noise components. Bias can be compensated for by adaptation and, in tasks in which the magnitude of noise varies across the environment, noise can be reduced by identifying and then acting in less noisy regions of the environment. Here we examine how these two processes interact when participants reach under a combination of an externally imposed visuomotor bias and noise. In a center-out reaching task, participants experienced noise (zero-mean random visuomotor rotations) that was target-direction dependent with a standard deviation that increased linearly from a least-noisy direction. They also experienced a constant bias, a visuomotor rotation that varied (across groups) from 0 to 40 degrees. Critically, on each trial, participants could select one of three targets to reach to, thereby allowing them to potentially select targets close to the least-noisy direction. The group who experienced no bias (0 degrees) quickly learned to select targets close to the least-noisy direction. However, groups who experienced a bias often failed to identify the least-noisy direction, even though they did partially adapt to the bias. When noise was introduced after participants experienced and adapted to a 40 degrees bias (without noise) in all directions, they exhibited an improved ability to find the least-noisy direction. We developed two models-one for reach adaptation and one for target selection-that could explain participants' adaptation and target-selection behavior. Our data and simulations indicate that there is a trade-off between adaptation and selection. Specifically, because bias learning is local, participants can improve performance, through adaptation, by always selecting targets that are closest to a chosen direction. However, this comes at the expense of improving performance, through selection, by reaching toward targets in different directions to find the least-noisy direction.

Object weight can be rapidly predicted, with low cognitive load, by exploiting learned associations between the weights and locations of objects.

Weight prediction is critical for dexterous object manipulation. Previous work has focused on lifting objects presented in isolation and has examined how the visual appearance of an object is used to predict its weight. Here we tested the novel hypothesis that when interacting with multiple objects, as is common in everyday tasks, people exploit the locations of objects to directly predict their weights, bypassing slower and more demanding processing of visual properties to predict weight. Using a three-dimensional robotic and virtual reality system, we developed a task in which participants were presented with a set of objects. In each trial a randomly chosen object translated onto the participant's hand and they had to anticipate the object's weight by generating an equivalent upward force. Across conditions we could control whether the visual appearance and/or location of the objects were informative as to their weight. Using this task, and a set of analogous web-based experiments, we show that when location information was predictive of the objects' weights participants used this information to achieve faster prediction than observed when prediction is based on visual appearance. We suggest that by "caching" associations between locations and weights, the sensorimotor system can speed prediction while also lowering working memory demands involved in predicting weight from object visual properties.NEW & NOTEWORTHY We use a novel object support task using a three-dimensional robotic interface and virtual reality system to provide evidence that the locations of objects are used to predict their weights. Using location information, rather than the visual appearance of the objects, supports fast prediction, thereby avoiding processes that can be demanding on working memory.

Decision-making in sensorimotor control.

Skilled sensorimotor interactions with the world result from a series of decision-making processes that determine, on the basis of information extracted during the unfolding sequence of events, which movements to make and when and how to make them. Despite this inherent link between decision-making and sensorimotor control, research into each of these two areas has largely evolved in isolation, and it is only fairly recently that researchers have begun investigating how they interact and, together, influence behaviour. Here, we review recent behavioural, neurophysiological and computational research that highlights the role of decision-making processes in the selection, planning and control of goal-directed movements in humans and nonhuman primates.

Human decision making anticipates future performance in motor learning.

It is well-established that people can factor into account the distribution of their errors in motor performance so as to optimize reward. Here we asked whether, in the context of motor learning where errors decrease across trials, people take into account their future, improved performance so as to make optimal decisions to maximize reward. One group of participants performed a virtual throwing task in which, periodically, they were given the opportunity to select from a set of smaller targets of increasing value. A second group of participants performed a reaching task under a visuomotor rotation in which, after performing a initial set of trials, they selected a reward structure (ratio of points for target hits and misses) for different exploitation horizons (i.e., numbers of trials they might be asked to perform). Because movement errors decreased exponentially across trials in both learning tasks, optimal target selection (task 1) and optimal reward structure selection (task 2) required taking into account future performance. The results from both tasks indicate that people anticipate their future motor performance so as to make decisions that will improve their expected future reward.

Rapid Visuomotor Responses Reflect Value-Based Decisions.

Cognitive decision-making is known to be sensitive to the values of potential options, which are the probability and size of rewards associated with different choices. Here, we examine whether rapid motor responses to perturbations of visual feedback about movement, which mediate low-level and involuntary feedback control loops, reflect computations associated with high-level value-based decision-making. In three experiments involving human participants, we varied the value associated with different potential targets for reaching movements by controlling the distributions of rewards across the targets (Experiment 1), the probability with which each target could be specified (Experiment 2), or both (Experiment 3). We found that the size of rapid and involuntary feedback responses to movement perturbations was strongly influenced by the relative value between targets. A statistical model of relative value that includes a term for risk sensitivity provided the best fit to the visuomotor response data, illustrating that feedback control policies are biased to favor more frequent task success at the expense of the overall extrinsic reward accumulated through movement. Importantly however, the regulation of rapid feedback responses was associated with successful pursuit of high-value task outcomes. This implies that when we move, the brain specifies a set of feedback control gains that enable low-level motor areas not only to generate efficient and accurate movement, but also to rapidly and adaptively respond to evolving sensory information in a manner consistent with value-based decision-making.SIGNIFICANCE STATEMENT Current theories of sensorimotor control suggest that, rather than selecting and planning the details of movements in advance, the role of the brain is to set time-varying feedback gains that continuously transform sensory information into motor commands by feedback control. Here, we examine whether the fastest motor responses to perturbations of movement, which mediate low-level and involuntary feedback control loops (i.e., reflexes), reflect computations associated with high-level, value-based decision-making. We find that rapid feedback responses during reaching reflect the relative probabilities and rewards associated with target options. This suggests that low-order components of the sensorimotor control hierarchy, which generate rapid and automatic responses, can continuously evaluate evolving sensory evidence and initiate responses according to the prospect of reward.

Motor memories in manipulation tasks are linked to contact goals between objects.

Skillful manipulation requires forming memories of object dynamics, linking applied force to motion. Although it has been assumed that such memories are linked to objects, a recent study showed that people can form separate memories when these are linked to different controlled points on an object (Heald JB, Ingram JN, Flanagan JR, Wolpert DM. Nat Hum Behav 2: 300-311, 2018). In that study, participants controlled the handle of a robotic device to move a virtual bar with circles (control points) on the left and right sides. Participants were instructed to move either the left or right control point to a target on the left or right, respectively, such that the required movement was constant. When these control points were paired with opposing force fields, adaptation was observed. In this previous study, both the controlled point and the target changed between contexts. To assess which of these factors is critical for learning, here, we used a similar paradigm but with a bar that automatically rotated as it was moved. In the first experiment, the bar rotated, such that the left and right control points moved to a common target. In the second experiment, the bar rotated such that a single control point moved to a target located on either the left or right. In both experiments, participants were able to learn opposing force fields applied in the two contexts. We conclude that separate memories of dynamics can be formed for different "contact goals," involving a unique combination of the controlled point on an object and the target location this point "contacts."NEW & NOTEWORTHY Skilled manipulation requires forming memories of object dynamics, previously assumed to be associated with entire objects. However, we recently demonstrated that people can form multiple motor memories when explicitly instructed to move different locations on an object to different targets. Here, we show that separate motor memories can be learned for different contact goals, which involve a unique combination of a control point and target.

Model-Free Robust Optimal Feedback Mechanisms of Biological Motor Control.

Sensorimotor tasks that humans perform are often affected by different sources of uncertainty. Nevertheless, the central nervous system (CNS) can gracefully coordinate our movements. Most learning frameworks rely on the internal model principle, which requires a precise internal representation in the CNS to predict the outcomes of our motor commands. However, learning a perfect internal model in a complex environment over a short period of time is a nontrivial problem. Indeed, achieving proficient motor skills may require years of training for some difficult tasks. Internal models alone may not be adequate to explain the motor adaptation behavior during the early phase of learning. Recent studies investigating the active regulation of motor variability, the presence of suboptimal inference, and model-free learning have challenged some of the traditional viewpoints on the sensorimotor learning mechanism. As a result, it may be necessary to develop a computational framework that can account for these new phenomena. Here, we develop a novel theory of motor learning, based on model-free adaptive optimal control, which can bypass some of the difficulties in existing theories. This new theory is based on our recently developed adaptive dynamic programming (ADP) and robust ADP (RADP) methods and is especially useful for accounting for motor learning behavior when an internal model is inaccurate or unavailable. Our preliminary computational results are in line with experimental observations reported in the literature and can account for some phenomena that are inexplicable using existing models.

Separate motor memories are formed when controlling different implicitly specified locations on a tool.

Skillful manipulation requires forming and recalling memories of the dynamics of objects linking applied force to motion. It has been assumed that such memories are associated with entire objects. However, we often control different locations on an object, and these locations may be associated with different dynamics. We have previously demonstrated that multiple memories can be formed when participants are explicitly instructed to control different visual points marked on an object. A key question is whether this novel finding generalizes to more natural situations in which control points are implicitly defined by the task. To answer this question, we used objects with no explicit control points and tasks designed to encourage the use of distinct implicit control points. Participants moved a handle, attached to a robotic interface, to control the position of a rectangular object ("eraser") in the horizontal plane. Participants were required to move the eraser straight ahead to wipe away a column of dots ("dust"), located to either the left or right. We found that participants adapted to opposing dynamics when linked to the left and right dust locations, even though the movements required for these two contexts were the same. Control conditions showed this learning could not be accounted for by contextual cues or the fact that the task goal required moving in a straight line. These results suggest that people naturally control different locations on manipulated objects depending on the task context and that doing so affords the formation of separate motor memories. NEW & NOTEWORTHY Skilled manipulation requires forming motor memories of object dynamics, which have been assumed to be associated with entire objects. However, we recently demonstrated that people can form multiple memories when explicitly instructed to control different visual points on an object. In this article we show that this novel finding generalizes to more natural situations in which control points are implicitly defined by the task.

An error-tuned model for sensorimotor learning.

Current models of sensorimotor control posit that motor commands are generated by combining multiple modules which may consist of internal models, motor primitives or motor synergies. The mechanisms which select modules based on task requirements and modify their output during learning are therefore critical to our understanding of sensorimotor control. Here we develop a novel modular architecture for multi-dimensional tasks in which a set of fixed primitives are each able to compensate for errors in a single direction in the task space. The contribution of the primitives to the motor output is determined by both top-down contextual information and bottom-up error information. We implement this model for a task in which subjects learn to manipulate a dynamic object whose orientation can vary. In the model, visual information regarding the context (the orientation of the object) allows the appropriate primitives to be engaged. This top-down module selection is implemented by a Gaussian function tuned for the visual orientation of the object. Second, each module's contribution adapts across trials in proportion to its ability to decrease the current kinematic error. Specifically, adaptation is implemented by cosine tuning of primitives to the current direction of the error, which we show to be theoretically optimal for reducing error. This error-tuned model makes two novel predictions. First, interference should occur between alternating dynamics only when the kinematic errors associated with each oppose one another. In contrast, dynamics which lead to orthogonal errors should not interfere. Second, kinematic errors alone should be sufficient to engage the appropriate modules, even in the absence of contextual information normally provided by vision. We confirm both these predictions experimentally and show that the model can also account for data from previous experiments. Our results suggest that two interacting processes account for module selection during sensorimotor control and learning.

Rapid target foraging with reach or gaze: The hand looks further ahead than the eye.

Real-world tasks typically consist of a series of target-directed actions and often require choices about which targets to act on and in what order. Such choice behavior can be assessed from an optimal foraging perspective whereby target selection is shaped by a balance between rewards and costs. Here we evaluated such decision-making in a rapid movement foraging task. On a given trial, participants were presented with 15 targets of varying size and value and were instructed to harvest as much reward as possible by either moving a handle to the targets (hand task) or by briefly fixating them (eye task). The short trial duration enabled participants to harvest about half the targets, ensuring that total reward was due to choice behavior. We developed a probabilistic model to predict target-by-target harvesting choices that considered the rewards and movement-related costs (i.e., target distance and size) associated with the current target as well as future targets. In the hand task, in comparison to the eye task, target choice was more strongly influenced by movement-related costs and took into account a greater number of future targets, consistent with the greater costs associated with arm movement. In both tasks, participants exhibited near-optimal behaviour and in a constrained version of the hand task in which choices could only be based on target positions, participants consistently chose among the shortest movement paths. Our results demonstrate that people can rapidly and effectively integrate values and movement-related costs associated with current and future targets when sequentially harvesting targets.

Rapid visuomotor feedback gains are tuned to the task dynamics.

Adaptation to novel dynamics requires learning a motor memory, or a new pattern of predictive feedforward motor commands. Recently, we demonstrated the upregulation of rapid visuomotor feedback gains early in curl force field learning, which decrease once a predictive motor memory is learned. However, even after learning is complete, these feedback gains are higher than those observed in the null field trials. Interestingly, these upregulated feedback gains in the curl field were not observed in a constant force field. Therefore, we suggest that adaptation also involves selectively tuning the feedback sensitivity of the sensorimotor control system to the environment. Here, we test this hypothesis by measuring the rapid visuomotor feedback gains after subjects adapt to a variety of novel dynamics generated by a robotic manipulandum in three experiments. To probe the feedback gains, we measured the magnitude of the motor response to rapid shifts in the visual location of the hand during reaching. While the feedback gain magnitude remained similar over a larger than a fourfold increase in constant background load, the feedback gains scaled with increasing lateral resistance and increasing instability. The third experiment demonstrated that the feedback gains could also be independently tuned to perturbations to the left and right, depending on the lateral resistance, demonstrating the fractionation of feedback gains to environmental dynamics. Our results show that the sensorimotor control system regulates the gain of the feedback system as part of the adaptation process to novel dynamics, appropriately tuning them to the environment.NEW & NOTEWORTHY Here, we test whether rapid visuomotor feedback responses are selectively tuned to the task dynamics. The responses do not exhibit gain scaling, but they do vary with the level and stability of task dynamics. Moreover, these feedback gains are independently tuned to perturbations to the left and right, depending on these dynamics. Our results demonstrate that the sensorimotor control system regulates the feedback gain as part of the adaptation process, tuning them appropriately to the environment.

When Optimal Feedback Control Is Not Enough: Feedforward Strategies Are Required for Optimal Control with Active Sensing.

Movement planning is thought to be primarily determined by motor costs such as inaccuracy and effort. Solving for the optimal plan that minimizes these costs typically leads to specifying a time-varying feedback controller which both generates the movement and can optimally correct for errors that arise within a movement. However, the quality of the sensory feedback during a movement can depend substantially on the generated movement. We show that by incorporating such state-dependent sensory feedback, the optimal solution incorporates active sensing and is no longer a pure feedback process but includes a significant feedforward component. To examine whether people take into account such state-dependency in sensory feedback we asked people to make movements in which we controlled the reliability of sensory feedback. We made the visibility of the hand state-dependent, such that the visibility was proportional to the component of hand velocity in a particular direction. Subjects gradually adapted to such a sensory perturbation by making curved hand movements. In particular, they appeared to control the late visibility of the movement matching predictions of the optimal controller with state-dependent sensory noise. Our results show that trajectory planning is not only sensitive to motor costs but takes sensory costs into account and argues for optimal control of movement in which feedforward commands can play a significant role.

Effective reinforcement learning following cerebellar damage requires a balance between exploration and motor noise.

Reinforcement and error-based processes are essential for motor learning, with the cerebellum thought to be required only for the error-based mechanism. Here we examined learning and retention of a reaching skill under both processes. Control subjects learned similarly from reinforcement and error-based feedback, but showed much better retention under reinforcement. To apply reinforcement to cerebellar patients, we developed a closed-loop reinforcement schedule in which task difficulty was controlled based on recent performance. This schedule produced substantial learning in cerebellar patients and controls. Cerebellar patients varied in their learning under reinforcement but fully retained what was learned. In contrast, they showed complete lack of retention in error-based learning. We developed a mechanistic model of the reinforcement task and found that learning depended on a balance between exploration variability and motor noise. While the cerebellar and control groups had similar exploration variability, the patients had greater motor noise and hence learned less. Our results suggest that cerebellar damage indirectly impairs reinforcement learning by increasing motor noise, but does not interfere with the reinforcement mechanism itself. Therefore, reinforcement can be used to learn and retain novel skills, but optimal reinforcement learning requires a balance between exploration variability and motor noise.

Rapid Visuomotor Corrective Responses during Transport of Hand-Held Objects Incorporate Novel Object Dynamics.

Numerous studies have shown that people are adept at learning novel object dynamics, linking applied force and motion, when performing reaching movements with hand-held objects. Here we investigated whether the control of rapid corrective arm responses, elicited in response to visual perturbations, has access to such newly acquired knowledge of object dynamics. Participants first learned to make reaching movements while grasping an object subjected to complex load forces that depended on the distance and angle of the hand from the start position. During a subsequent test phase, we examined grip and load force coordination during corrective arm movements elicited (within ∼150 ms) in response to viewed sudden lateral shifts (1.5 cm) in target or object position. We hypothesized that, if knowledge of object dynamics is incorporated in the control of the corrective responses, grip force changes would anticipate the unusual load force changes associated with the corrective arm movements so as to support grasp stability. Indeed, we found that the participants generated grip force adjustments tightly coupled, both spatially and temporally, to the load force changes associated with the arm movement corrections. We submit that recently learned novel object dynamics are effectively integrated into sensorimotor control policies that support rapid visually driven arm corrective actions during transport of hand held objects. Significance statement: Previous studies have demonstrated that the motor system can learn, and make use of, internal models of object dynamics to generate feedforward motor commands. However, it is not known whether such internal models are incorporated into rapid, automatic arm movement corrections that compensate for errors that arise during movement. Here we demonstrate, for the first time, that internal models of novel object dynamics are integrated into rapid corrective arm movements made in response to visuomotor perturbations that, importantly, do not directly perturb the object.

The sequential encoding of competing action goals involves dynamic restructuring of motor plans in working memory.

Recent neural and behavioral findings provide support for the influential idea that in situations in which multiple action options are presented simultaneously, we prepare action plans for each competing option before deciding between and executing one of those plans. However, in natural, everyday environments, our available action options frequently change from one moment to the next, and there is often uncertainty as to whether additional options will become available before having to select a particular course of action. Here, with the use of a target-directed reaching task, we show that in this situation, the brain specifies a competing action for each new, sequentially presented potential target and that recently formed action plans can be revisited and updated so as to conform with separate, more newly developed, plans. These findings indicate that the brain forms labile motor plans for sequentially arising target options that can be flexibly restructured to accommodate new motor plans.

Enhanced crosslimb transfer of force-field learning for dynamics that are identical in extrinsic and joint-based coordinates for both limbs.

Humans are able to adapt their motor commands to make accurate movements in novel sensorimotor environments, such as when wielding tools that alter limb dynamics. However, it is unclear to what extent sensorimotor representations, obtained through experience with one limb, are available to the opposite, untrained limb and in which form they are available. Here, we compared crosslimb transfer of force-field compensation after participants adapted to a velocity-dependent curl field, oriented either in the sagittal or the transverse plane. Due to the mirror symmetry of the limbs, the force field had identical effects for both limbs in joint and extrinsic coordinates in the sagittal plane but conflicting joint-based effects in the transverse plane. The degree of force-field compensation exhibited by the opposite arm in probe trials immediately after initial learning was significantly greater after sagittal (26 ± 5%) than transverse plane adaptation (9 ± 4%; P < 0.001), irrespective of whether participants learned initially with the left or the right arm or via abrupt or gradual exposure to the force field. Thus transfer was impaired when the orientation of imposed dynamics conflicted in intrinsic coordinates for the two limbs. The data reveal that neural representations of novel dynamics are only partially available to the opposite limb, since transfer is incomplete even when force-field perturbation is spatially compatible for the two limbs, according to both intrinsic and extrinsic coordinates.

Principles of sensorimotor learning.

The exploits of Martina Navratilova and Roger Federer represent the pinnacle of motor learning. However, when considering the range and complexity of the processes that are involved in motor learning, even the mere mortals among us exhibit abilities that are impressive. We exercise these abilities when taking up new activities - whether it is snowboarding or ballroom dancing - but also engage in substantial motor learning on a daily basis as we adapt to changes in our environment, manipulate new objects and refine existing skills. Here we review recent research in human motor learning with an emphasis on the computational mechanisms that are involved.