Abstract:
The state estimation (SE) theory is the modern instantiation of a classic sensorimotor theory called the reafferent principle (Von Holst and Mittelstaedt, 1950). It estimates the body's or its limbs' kinematic state based on optimal integration of the internal model's predictive signals (based on an efference copy and other contextual signals) and external sensory signals. It generates the so-called sensory prediction error that can be used to estimate the state, optimize movement and perception, as well as to train the internal model. State estimation gives rise to the sensory prediction that is employed to cancel the reafference inputs or the sensory consequences of self-motion to enhance the information from more relevant stimuli. Sensory gating (SG), a process that modifies the sensory signal as well, attenuates external sensory responses that occur during the movements (Hentschke et al. (2006), (Seki et al., 2003, Chapman et al., 1987b, Chapman et al., 1988, Rushton et al., 1981, Ghez and Pisa, 1972, Chapman et al., 1987a). It is currently not known whether attenuation caused by SE and SG is the expression of the same or different functional (behavioral) systems, and whether it is based on the same or different neuronal circuits. In fact, many studies studying SG discussed their results in the framework of SE. Others that meant to study SE show phenomena better assigned to SG.
Here, I established an experimental paradigm in mice with the aim to separate the two processes. To this end I used and extended the open loop approach first pioneered by Curtis Bell (1982) in weakly electric fish. This method records the neuronal motor command while blocking its motor outcome. Introduction of an artificial sensory consequence after the onset of a motor command then allows to probe the neuronal prediction signal by omitting the sensory consequence. My first extension was to realize the open loop approach in a mammal - in the tactile whisker-related system of head-fixed mice, trained to generate a whisker reaching movement. This was required, as SG has been described so far only in mammals. The second extension was to add rare test stimuli that deliberately varied the delay of a sensory consequence after learning the sensory consequence at a fixed delay. It turned out that SG is active throughout the movement while SE is active only at the time point of the predicted sensory consequence, a major difference between the processes, and the reason for my success to separate them.
The experimental setup included the chronic implantation of a micro-electrode in the facial nucleus to extracellularly measure the whisker motor command, and multi-electrode devices in the somatosensory cortex to record the tactile sensory signals. The reafferent loop was opened by surgically disrupting two branches of the distal facial motor nerve, which innervate the intrinsic whisker muscles, and thus, paralyzing whisker movements. Artificial sensory consequences of an intended reach were realized by deflecting the immobilized whisker using a Piezo actuator. In a different, closed-loop approach, I left the movement intact and electrically stimulated the trigeminal nucleus to mimic the sensory consequence of the intended whisker movement. The artificial sensory consequence was presented at the trained (predicted) delay. In rare test trials it was shifted to other times during the movement, to times between intended movements, or was omitted. The tactile responses were recorded in the primary somatosensory cortex (S1) and consistently were found to be strongest in between movements, significantly attenuated at a medium level with shifted stimuli, and attenuated significantly stronger at predicted delays (12 mice studied in open loop and 3 mice studied in closed loop). Somatosensory attenuation due to SE turned out to be adaptable to a new delay within a few hundred trials and acting at a temporal precision of tens of milliseconds. It could be trained at delays up to 200-300ms from motor command onset. In contrast, SG was observed independently of learning the sensory consequence. It was present at all times up to 500ms after the motor command onset. Using a linear array of 16 electrodes distributed across the six neocortical layers, I could show that the neuronal reflections of SE and SG are distributed across all layers of S1.
The significance of my work is that for the first time two neuronal functional systems could be disentangled that act as movement-dependent attenuators of sensory signal flow. SE is highly likely predictive of detailed sensory consequences of movement. One speculation close at hand is that it may be dependent on cerebellar function. Whether SG-mediated attenuation is of predictive nature needs to be studied in the future. In view of its known dependence on neocortical circuits (Chakrabarti and Schwarz, 2018), it may well be related to higher functions like attentional processes or reward predictions.
State estimation