Maurice A. Smith

Short sub-100 ms visual feedback latencies are common in many types of human-computer interactions yet are known to markedly reduce performance in a wide variety of motor tasks from simple pointing to operating surgical robotics. It remains unclear, however, whether these latencies impair not only skilled motor performance but also the implicit sensorimotor learning that underlies its acquisition. Inspired by neurophysiological findings showing that cerebellar LTD and cortical LTP would both be disrupted by sub-100 ms latencies, we hypothesized that implicit sensorimotor learning may be particularly sensitive to these short latencies. Remarkably, we find that improving latency by just 60 ms, from 85 to 25 ms in continuous-feedback experiments, increases implicit learning by 50% and proportionally decreases explicit learning. This resulted in a dramatic reorganization of sensorimotor memory from a 45/55 to a 70/30 implicit/explicit ratio. This 70/30 ratio is more than double that observed in any previous study examining the effect of latency on sensorimotor learning, including a recent study which provided time-advanced visual feedback, suggesting that the low-latency continuous visual feedback we provided is critical for efficiently driving implicit learning. We go on to show that implicit sensorimotor learning is considerably more sensitive to latencies in the sub-100 ms range than to higher latencies, in line with the latency-specific neural plasticity that has been observed. This suggests a clear benefit for latency reduction in computer-based training that involves implicit sensorimotor learning and that across-study differences in computer-based experiments that have examined implicit sensorimotor learning might be explained by differences in unmeasured feedback latencies.

The cerebellum is critical for sensorimotor learning. The specific contribution that it makes, however, remains unclear. Inspired by the classic finding that for declarative memories, medial temporal lobe (MTL) structures provide a gateway to the formation of long-term memory but are not required for short-term memory, we hypothesized that for sensorimotor memories, the cerebellum may play an analogous role. Here, we studied the sensorimotor learning of individuals with severe ataxia from cerebellar degeneration. We dissected the memories they formed during sensorimotor learning into a short-term temporally-volatile component, that decays rapidly with a time constant of just 15 to 20 s and thus cannot lead to long-term retention, and a longer-term temporally-persistent component that is stable for 60 s or more and leads to long-term retention. Remarkably, we find that these individuals display dramatically reduced levels of temporally-persistent sensorimotor memory, despite spared and even elevated levels of temporally-volatile sensorimotor memory. In particular, we find both impairment that systematically worsens with memory window duration over shorter memory windows (<12 s) and near-complete impairment of memory maintenance over longer memory windows (>25 s). This dissociation uncovers a unique role for the cerebellum as a gateway for the formation of long-term but not short-term sensorimotor memories, mirroring the role of the MTL for declarative memories. It thus reveals the existence of distinct neural substrates for short-term and long-term sensorimotor memory, and it explains both the trial-to-trial differences identified in this study and long-standing study-to-study differences in the effects of cerebellar damage on sensorimotor learning ability.

Sensorimotor learning can change the tuning of neurons in motor-related brain areas and rotate their preferred directions (PDs). These PD rotations are commonly interpreted as reflecting motor command changes; however, cortical neurons that display PD rotations also contribute to sensorimotor learning. Sensorimotor learning should, therefore, alter not only motor commands but also the tuning of neurons responsible for this learning, and thus impact subsequent learning ability. Here, we investigate this possibility with computational modeling and by directly measuring adaptive responses during sensorimotor learning in humans. Modeling shows that the PD rotations induced by sensorimotor learning, predict specific anisotropic changes in PD distributions that in turn predict a specific spatial pattern of changes in learning ability. Remarkably, experiments in humans then reveal large, systematic changes in learning ability in a spatial pattern that precisely reflects these model-predicted changes. We find that this pattern defies conventional wisdom and implements Newton's method, a learning rule where the step size is inversely proportional rather than proportional to the learning gradient's amplitude, limiting overshooting in the adaptive response. Our findings indicate that PD rotation provides a mechanism whereby the motor system can simultaneously learn how to move and learn how to learn.

Short sub-100ms visual feedback latencies are common in many types of human-computer interactions yet are known to markedly reduce performance in a wide variety of motor tasks from simple pointing to operating surgical robotics. These latencies are also present in the computer-based experiments used to study the sensorimotor learning that underlies the acquisition of motor performance. Inspired by neurophysiological findings showing that cerebellar LTD and cortical LTP would both be disrupted by sub-100ms latencies, we hypothesized that implicit sensorimotor learning may be particularly sensitive to these short latencies. Remarkably, we find that improving latency by just 60ms, from 85 to 25ms in latency-optimized experiments, increases implicit learning by 50% and proportionally decreases explicit learning, resulting in a dramatic reorganization of sensorimotor memory. We go on to show that implicit sensorimotor learning is considerably more sensitive to latencies in the sub-100ms range than at higher latencies, in line with the latency-specific neural plasticity that has been observed. This suggests a clear benefit for latency reduction in computer-based training that involves implicit sensorimotor learning and that across-study differences in implicit motor learning might often be explained by disparities in feedback latency.

The cerebellum is critical for sensorimotor learning. The specific contribution that it makes, however, remains unclear. Inspired by the classic finding that, for declarative memories, medial temporal lobe structures provide a gateway to the formation of long-term memory but are not required for short-term memory, we hypothesized that, for sensorimotor memories, the cerebellum may play an analogous role. Here we studied the sensorimotor learning of individuals with severe ataxia from cerebellar degeneration. We dissected the memories they formed during sensorimotor learning into a short-term temporally-volatile component, that decays rapidly with a time constant of just 15-20sec and thus cannot lead to long-term retention, and a longer-term temporally-persistent component that is stable for 60 sec or more and leads to long-term retention. Remarkably, we find that these individuals display dramatically reduced levels of temporally-persistent sensorimotor memory, despite spared and even elevated levels of temporally-volatile sensorimotor memory. In particular, we find both impairment that systematically increases with memory window duration over shorter memory windows (<12 sec) and near-complete impairment of memory maintenance over longer memory windows (>25 sec). This dissociation uncovers a new role for the cerebellum as a gateway for the formation of long-term but not short-term sensorimotor memories, mirroring the role of the medial temporal lobe for declarative memories. It thus reveals the existence of distinct neural substrates for short-term and long-term sensorimotor memory, and it explains both newly-identified trial-to-trial differences and long-standing study-to-study differences in the effects of cerebellar damage on sensorimotor learning ability. Significance Statement: A key discovery about the neural underpinnings of memory, made more than half a century ago, is that long-term, but not short-term, memory formation depends on neural structures in the brain's medial temporal lobe (MTL). However, this dichotomy holds only for declarative memories - memories for explicit facts such as names and dates - as long-term procedural memories - memories for implicit knowledge such as sensorimotor skills - are largely unaffected even with substantial MTL damage. Here we demonstrate that the formation of long-term, but not short-term, sensorimotor memory depends on a neural structure known as the cerebellum, and we show that this finding explains the variability previously reported in the extent to which cerebellar damage affects sensorimotor learning.

Memories are easier to relearn than learn from scratch. This advantage, known as savings, has been widely assumed to result from the reemergence of stable long-term memories. In fact, the presence of savings has often been used as a marker for whether a memory has been consolidated. However, recent findings have demonstrated that motor learning rates can be systematically controlled, providing a mechanistic alternative to the reemergence of a stable long-term memory. Moreover, recent work has reported conflicting results about whether implicit contributions to savings in motor learning are present, absent, or inverted, suggesting a limited understanding of the underlying mechanisms. To elucidate these mechanisms, we investigate the relationship between savings and long-term memory by experimentally dissecting the underlying memories based on short-term (60-s) temporal persistence. Components of motor memory that are temporally-persistent at 60 s might go on to contribute to stable, consolidated long-term memory, whereas temporally-volatile components that have already decayed away by 60 s cannot. Surprisingly, we find that temporally-volatile implicit learning leads to savings, whereas temporally-persistent learning does not, but that temporally-persistent learning leads to long-term memory at 24 h, whereas temporally-volatile learning does not. This double dissociation between the mechanisms for savings and long-term memory formation challenges widespread assumptions about the connection between savings and memory consolidation. Moreover, we find that temporally-persistent implicit learning not only fails to contribute to savings, but also that it produces an opposite, anti-savings effect, and that the interplay between this temporally-persistent anti-savings and temporally-volatile savings provides an explanation for several seemingly conflicting recent reports about whether implicit contributions to savings are present, absent, or inverted. Finally, the learning curves we observed for the acquisition of temporally-volatile and temporally-persistent implicit memories demonstrate the coexistence of implicit memories with distinct time courses, challenging the assertion that models of context-based learning and estimation should supplant models of adaptive processes with different learning rates. Together, these findings provide new insight into the mechanisms for savings and long-term memory formation.

Data for "A Double Dissociation between Savings and Long-Term Memory in Motor Learning"<br> <br> For analysis code and any updates, please visit https://github.com/AlkisMH/Savings_vs_Long_Term_Memory<br> <br> For any questions, please contact Alkis Hadjiosif (alkis [at] seas [dot] harvard [dot] edu; ahadjiosif [at] gmail [dot] com) These .mat files contain data collected for Experiments 1-4 and S1, used to reproduce Figures 1-6 and S1 in the paper. Data for each experiment are in the form of a Matlab structure, with each field consisting of a [Num_trials x Num_participants] matrix. The fields are: <strong>TN:</strong> Trial Number <strong>VF:</strong> Whether visual feedback was given during the trial (1: online visual feedback; 2: no visual feedback) <strong>Rotation:</strong> The visuomotor rotation imposed on each trial (in degrees). CCW is positive, CW is negative. <strong>Wait:</strong> Whether, right before the trial, a wait time was imposed (1) or not (0). Trials immediately following breaks are indicated by (2). <strong>Instruction:</strong> Whether an instruction was given for the trial (1) or not (0). Only present in Experiment 3. Note that Experiment 3 includes: -&gt; Two specific instruction trials during learning and and two during relearning, before and after the 1-minute wait ("Move your hand to the center of the target","Move your hand to the far end of the target") -&gt; Six random instruction trials before initial learning and six before relearning ("Move your hand to the left/right/near/far end of the target"), to familiarize participants with the instruction process <strong>ITI:</strong> Time (in seconds) since the last trial (regardless of direction). <strong>theta_target:</strong> Target direction (in degrees) relative to the 12 o'clock position. Only included for Experiment 4: target is always at 12 o'clock for Experiments 1-3. <strong>theta:</strong> = reaching direction relative to theta_target, measured 150ms into the movement. Note that this is flipped based on the rotation sign so that adaptation towards the imposed visuomotor rotation is always positive. <strong>theta_end:</strong> = reaching direction relative to theta_target, measured at the end of movement. Only included for Experiment 4: it is to be used for the no visual feedback blocks in Experiment 4. Note for <strong>Experiment 4</strong>: the last two blocks (last 114 trials) were done on day 2.

Data for "A Double Dissociation between Savings and Long-Term Memory in Motor Learning"<br> <br> For analysis code and any updates, please visit https://github.com/AlkisMH/Savings_vs_Long_Term_Memory<br> <br> For any questions, please contact Alkis Hadjiosif (alkis [at] seas [dot] harvard [dot] edu; ahadjiosif [at] gmail [dot] com) These .mat files contain data collected for Experiments 1-4 and S1, used to reproduce Figures 1-6 and S1 in the paper. Data for each experiment are in the form of a Matlab structure, with each field consisting of a [Num_trials x Num_participants] matrix. The fields are: <strong>TN:</strong> Trial Number <strong>VF:</strong> Whether visual feedback was given during the trial (1: online visual feedback; 2: no visual feedback) <strong>Rotation:</strong> The visuomotor rotation imposed on each trial (in degrees). CCW is positive, CW is negative. <strong>Wait:</strong> Whether, right before the trial, a wait time was imposed (1) or not (0). Trials immediately following breaks are indicated by (2). <strong>Instruction:</strong> Whether an instruction was given for the trial (1) or not (0). Only present in Experiment 3. Note that Experiment 3 includes: -&gt; Two specific instruction trials during learning and and two during relearning, before and after the 1-minute wait ("Move your hand to the center of the target","Move your hand to the far end of the target") -&gt; Six random instruction trials before initial learning and six before relearning ("Move your hand to the left/right/near/far end of the target"), to familiarize participants with the instruction process <strong>ITI:</strong> Time (in seconds) since the last trial (regardless of direction). <strong>theta_target:</strong> Target direction (in degrees) relative to the 12 o'clock position. Only included for Experiment 4: target is always at 12 o'clock for Experiments 1-3. <strong>theta:</strong> = reaching direction relative to theta_target, measured 150ms into the movement. Note that this is flipped based on the rotation sign so that adaptation towards the imposed visuomotor rotation is always positive. <strong>theta_end:</strong> = reaching direction relative to theta_target, measured at the end of movement. Only included for Experiment 4: it is to be used for the no visual feedback blocks in Experiment 4. Note for <strong>Experiment 4</strong>: the last two blocks (last 114 trials) were done on day 2.