Motor control: Neural correlates of optimal feedback control theory

Recent work is revealing neural correlates of a leading theory motor control. By linking an elegant series behavioral experiments with inactivation in macaques computational models, new study shows that premotor and parietal areas can be mapped onto model for optimal feedback We are constantly adapting. Whether it to shoes, or walking on icy terrain, we have the capacity continually update our actions goal-directed manner. The how do this — control (OFC) has been upheld by impressive number studies humans other animals over last 20 years, yet circuits implement such continue remain elusive. Now, reported issue Current Biology, Takei et al.1Takei T. Lomber S.G. Cook D.J. Scott S.H. Transient deactivation dorsal cortex area 5 impairs limb macaques.Curr. Biol. 2021; 31: 1476-1487Abstract Full Text PDF PubMed Scopus (21) Google Scholar chipped away at long-standing question providing important clues may begin theory. origins date back nearly 300 but 1950s Wiener’s cybernetics movement brought forth notion intelligent behavior rooted OFC burst scene 2002 seminal Todorov Jordan2Todorov E. Jordan M.I. Optimal as coordination.Nat. Neurosci. 2002; 5: 1226-1235Crossref (2024) Scholar. They postulated that, deploying stochastic control, system would only correct movements task-relevant dimensions. This allows variability task-irrelevant ones, ‘minimum intervention’ principle nicely explains many behaviors humans. was breakthrough several respects. Firstly, framework requires multiple parts: internal forward enables computation signals (given noisy, delayed feedback), which integrated knowledge body dynamics available copy out-going command (the so-called efference copy; Figure 1A) via state-estimator. Secondly, nature corrections beautifully matched observations. allowed exploit redundancy improve performance, matches observations variability, occur mostly closest end-goal (for example, correcting hand mid-trajectory match average trajectory template not required; what required hitting end-goal). Lastly, contrast theories active inference3Friston K. Mattout J. Kilner Action understanding inference.Biol. Cybernet. 2011; 104: 137-160Crossref (394) error learning4Kawato M. Gomi H. A four regions cerebellum based feedback-error learning.Biol. 1992; 68: 95-103Crossref (401) Scholar, does require explicitly considers noise delays, constraints biological systems. Shortly thereafter, Scott5Scott basis volitional control.Nat. Rev. 2004; 532-545Crossref (624) hypothesized because so powerful explaining human consists elements, could highly valuable mapping (volitional) He laid out crucial For observed tuning-properties (M1) neurons activity were found change behaviorally dependent idea spawned years research into without doubt must include notions models circuits. Techniques optogenetics chemogenetics allow experimentalists design where one spatially (in brain) temporally space) perturb during behaviors6Boyden E.S. Zhang F. Bamberg Nagel G. Deisseroth Millisecond-timescale, genetically targeted optical activity.Nat. 2005; 8: 1263-1268Crossref (3439) 7Sternson S.M. Roth B.L. Chemogenetic tools interrogate brain functions.Annu. 2014; 37: 387-407Crossref (332) 8Azim Jiang Alstermark B. Jessell T.M. Skilled reaching relies V2a propriospinal circuit.Nature. 508: 357-363Crossref (205) 9Mathis M.W. Mathis A. Uchida N. Somatosensory plays essential role forelimb adaptation mice.Neuron. 2017; 93: 1493-1503Abstract (88) While limited rodent studies, future holds these types non-human primates. In their work, causally tested primate species use cooling probes macaque cortex1Takei Scholar,10Meyer-Lohmann Conrad Matsunami Brooks V.B. Effects dentate precentral unit following torque pulse injections elbow movements.Brain Res. 1975; 94: 237-251Crossref (62) prior colleagues11Omrani Murnaghan C.D. Pruszynski J.A. Distributed task-specific processing somatosensory voluntary control.eLife. 2016; 5e13141Crossref (54) showed respond rapidly perturbations. Specifically, they recorded five cortical involved control: (A5), primary 1 2 (S1, S2), M1, (PMd)). delays sensory information from periphery Namely, if limb-perturbations ongoing task, measured response times A5, S1 ∼25 milliseconds, target-selection first causes responses PMd M1 (with trailing behind). intriguing evidence areas, might play corretions. Building earlier work5Scott Scholar,11Omrani Scholar,12Pruszynski Kurtzer I. Nashed J.Y. Omrani Brouwer Primary underlies multi-joint integration fast control.Nature. 478: 387-390Crossref (216) now A5 perturbations applied postural stabilization task. First, paradigm performed ‘inactivation’ investigate parameters differential effects predicted behavior. key test: ‘Kalman gain’ (K), term provides uncertainty measure (that modulates learning rate); observation (Ĥ); ‘feedback’ (L) ‘state estimator’ estimated state body) controller region sends commands). Formally, al. test variations all three components model, estimator policy model. true x transformed measurements (observations) y means matrix H passed estimator. Based measurement noise, trades off current estimate, controlled Kalman gain K (Figure 1A). computes difference actual vs. expected transforming output Ĥ. estimation result finally fed policy, scaled L Modifying aforementioned non-trivially affect results parameter testing summarized 1B; namely, inactivate ‘L’ predict speed, strong reduction L, H, there also endpoint error. contrast, normal, modified, error, no response-speed changes. Next, used spatial brain10Meyer-Lohmann Scholar,13Lomber Payne B.R. Horel cryoloop: adaptable reversible method electrophysiological assessment function.J. Methods. 1999; 86: 179-194Crossref (175) PMd, suggesting different roles Cooling increased while inhibition both reduced accuracy, among metrics. former finding directly links effect downscaling highlighting input qualitatively 1). additional combining authors went show impairments errors speed add up linearly scale sublinear maximally deviation. simultaneous attenuation gains theory, input, reproduce 1B, grey box prediction). Adding link between gain, consider partial third smaller gain. yields characteristic setting: full experiment, impairment longer impacts (previously, increase plus experiments). Alternatively reducing single modeled increasing levels parameters, respectively. However, experiment solely explained Overall, build previous findings regarding specific predictions experimental domains. As summarize (L), (K, H), future, will address some limitations ways. demonstrate predictions, behavior, likely enriched taking account themselves. Fundamentally, phenomenological its unresolved. Moreover, numerous potential implementations circuit level, provide concrete hypotheses important. To direction, more fine-grained yield richer (and, course, cell-type-specific approaches become important). same applies deactivating individual regions. These choices influencing apply identify interesting working points unique patterns. Reproducing patterns adding them author’s combinatorial 1B) shed even light brains components, indicate complex needed accurate implementation14Shadmehr R. Krakauer neuroanatomy control.Exp. Brain 2008; 185: 359-381Crossref (725) Scholar,15Merel Botvinick Wayne Hierarchical mammals machines.Nat. Commun. 2019; 10: 5489Crossref (85) macaquesTakei al.Current BiologyFebruary 15, 2021In BriefTakei temporary distinctly arm monkeys, reflecting distinct aspects estimation, causal broad Full-Text Open Archive