Bio Mirror / Memory, Motor Control, and Embodiment

The Cerebellum and Internal Models

The clearest structural mirror in this entire section: a supervised learner computing what the inverse-Jacobian track computes.

The structure

The cerebellum holds more than half the neurons in the brain in a strikingly regular, almost crystalline circuit. Granule cells — the most numerous neuron type in the body — receive mossy fiber input and send parallel fibers across enormous numbers of Purkinje cells. Each Purkinje cell receives tens of thousands of parallel fiber synapses and exactly one climbing fiber from the inferior olive. That one-to-one climbing fiber is the anatomical detail that made the computational theory possible.

The function it dictates

The cerebellum learns internal models of the body and its interactions. A forward model predicts the sensory consequence of a motor command, which allows correction to begin before sensory feedback arrives — essential because that feedback is far too slow for fast movement. An inverse model computes the motor command required to achieve a desired outcome. The climbing fiber carries an error signal that drives long-term depression at the parallel fiber synapses, so the circuit is a supervised learner with an explicit teaching input. Cerebellar damage produces dysmetria — reaching that overshoots and undershoots with corrective oscillation, precisely the signature of a controller that has lost its internal model.

The mirror in AI

Supervised learning with an explicit error signal, and — since the cerebellum is also engaged in cognitive and language tasks, not only motor ones — a candidate for a general prediction-and-correction module operating alongside cortex. The Marr-Albus-Ito theory, formulated from anatomy in the late 1960s, is one of the few places where a learning rule was predicted from wiring diagram before it was measured.

The mirror in robotics

This is the mirror the embodied-control track was built for. A cerebellar inverse model does what inverse kinematics does: given a desired end state, produce the commands that achieve it. A forward model does what a state predictor in a Kalman filter or a model-predictive controller does: predict the consequence of a command so correction need not wait for feedback. Feedback-error learning — training an inverse model from the corrective output of a feedback controller — is an explicitly cerebellum-derived control architecture, and it works.

Where the analogy breaks

The cerebellum learns its models online, continuously, from ongoing experience, without a training phase and without catastrophically forgetting the models it already has. Robotic learned controllers are typically trained offline and then frozen, because online learning risks instability. The cerebellum also handles a body whose dynamics change — growth, fatigue, injury, a tool in the hand — by continuously re-fitting, which no deployed robot does well. And the biological error signal is sparse and delayed rather than a clean per-timestep gradient.

Open questions

How does the cerebellum learn online without destabilizing the controller it is part of?
What is the cerebellum computing during purely cognitive tasks — the same prediction, on different content?

In Kiru Lab

Derivatives and Gradients Inverse Jacobian Methods Redundancy and the Null Space Learned Controllers and the Gap to Reality