Cortex and cerebellum: two papers on motor learning and changing representations

Stable control under delayed feedback is hard, both in robotics and in neuroscience. Classical models of the cerebellum explain delay compensation through forward prediction. They do not account for the fast online corrections and equally fast adaptation seen in biological systems.

Prediction plus feedback

Ana Carolina Filipe, Rui Ponte Costa and Cláudia Soares propose a cerebellum-inspired control scheme. It combines multiplexed predictive representations with internal feedback. The encoder shares kinematic variables and task-relevant error signals, so the movement can be corrected precisely despite a delayed feedback signal. The authors show that adding feedback to the cerebellar loop alone clearly speeds up adaptation, cutting learning time by an order of magnitude. Under delay, however, predictions based on a single signal are not enough. Only multiplexing and feedback together give a consistent mechanism for simultaneous online control and fast learning.

Changing representations without stimulation

The second paper starts from a different question. In neuropsychiatry the goal is often not to read brain activity but to change it, for example to weaken a negative affective charge or an overly consolidated memory. Marco Rothernel's team proposes a machine-learning framework of neural surrogate models. It proposes candidate representation changes and tests the predicted perceptual effects from images of stimulus-evoked fMRI activity, without physical stimulation. The scheme combines fMRI decoding, deep generative modelling and control in latent space with constraints. Valence and memorability are only working examples here.

The material came from more than 36,000 image-fMRI observations from four deeply scanned participants in the Natural Scenes Dataset. Person-fitted models reproduced the coarse generative structure of visual cortex, reaching 0.79 to 0.88 in a two-way identification task against a chance level of 0.5. Graded perturbations were reconstructed as images and assessed with automated scorers and in human ratings from 7,200 trials performed by 18 participants. In the base model, valence shifted from -0.61 to +1.03 standard deviations, and memorability from -1.34 to +1.45 standard deviations. In the human ratings, valence shifted in the predicted direction, with a mean slope of 0.038 standard deviations per unit of coefficient, a confidence interval of 0.003 to 0.074, and a positive slope in 16 of 18 participants. Perceived memorability did not change reliably.

The authors highlight the limitations and do not hide them. Agreement with the automated scorer was only suggestive for valence and weak for memorability, and extreme perturbations drifted from the original stimulus, so the intended change has to be weighed against loss of fidelity. This is not a ready treatment, but a testable tool for designing and behaviourally testing candidate targets for future neuromodulation.