Skip to main content
Back to timeline
bioRxivSource publication:

A cerebellar-medial temporal lobe circuit learns only to predict the next state, yet infers context and transfers to new ones

Synopsis

A circuit assembled from known anatomy pairs a cerebellar forward model, which predicts the next state conditioned on a latent from the medial temporal lobe, with a medial temporal lobe module that updates that latent from the cerebellum's prediction error. Trained only on next-state prediction, it cut held-out velocity error across twelve uncued motor contexts from 0.0348 to 0.0138, below the 0.0223 of a context-free model, and organized contexts by family and parameter so a downstream motor cortex could retrieve old contexts and compensate for unseen ones on first reach.

Source-provided article image: Structure learning and context inference emerge from next-state prediction in a cerebellar-medial temporal lobe circuit
Figure 1 · bioRxiv · Page 3

Interpretation

A circuit built from a cerebellar forward model and a recurrent medial temporal lobe module, trained only to predict the next observation, learns to infer hidden context. Human motor learning is well described by context inference, but which brain circuit performs that inference was unknown; the cerebellum was thought to hold multiple internal models and switch among them by context, the medial temporal lobe to encode context during inference tasks, and the two regions are reciprocally connected. Across twelve uncued contexts, held-out velocity error fell from 0.0348±0.0006 to 0.0138±0.003; a context-free model averaging the twelve context-conditioned predictions reached 0.0223±0.00049, and a model without the medial temporal lobe reached 0.0228±0.0002.

Context information concentrates in the 32-dimensional latent read out from the medial temporal lobe and can be decoded linearly. The cerebellar module represents position and the medial temporal lobe's recurrent hidden state reflects heading and speed, so context does not appear directly in those signals. On held-out data, linear decoders identified the twelve contexts with 60.6±2.4% accuracy against 8.3% chance; zeroing the prediction error reduced accuracy by 36.7±3.7% relative, and shuffling it across trials reduced accuracy by 65.7±2.2%.

With the circuit frozen, the latent places trained and novel contexts in one structured space, with each perturbation family in its own subspace and ordered by parameter within it. Discrete context models treat each context as an independent entity and learn no relations between them, so they can retrieve but not transfer. Under 24 novel contexts, rotations lay along the first principal component by angle, curls along the second by signed gain, and drifts formed concentric rings in the third and fourth components encoding direction and speed; novel rotations beyond the training range extended the axis with compressed spacing.

A motor cortex module receiving this latent reproduces context-dependent retrieval, savings after washout, spontaneous recovery and evoked recovery, and compensates for an unseen perturbation on first reach. Motor cortex alone, or motor cortex plus cerebellum, has no context signal and can improve only through context-independent competence; the discrete context model can retrieve but not transfer. On returning to the −60° context, the circuit and the discrete context model showed protection indices of 1 [0.9-1.02] and 1 [0.9-1], against 0.7 [0.5-0.78] and 0.67 [0.44-0.78] for the other two circuits; on first reaches in a novel 60° context, only this circuit compensated, by 0.28 [0.19-0.34] of the ideal.

Perspective

The demonstrated scope is simulation: the circuit consists of a cerebellar forward model and a recurrent module standing for entorhinal cortex and medial temporal lobe, run on twelve motor contexts and several memory and cognitive tasks, and compared with motor cortex alone, motor cortex plus cerebellum, and a discrete context model. The authors explicitly propose next steps of testing new predictions, such as whether context inference in the serial reversal task is cerebellar-dependent and whether structural generalization across rotations or curls is medial temporal lobe-dependent.

The authors state that the circuit reproduces previously reported data and does not itself test new predictions; it is schematic, collapsing entorhinal cortex and hippocampus into one recurrent module, ignoring cerebellar fan-in fan-out connectivity, and assuming cerebellar and medial temporal lobe weights are frozen while motor cortex learns navigation. The work therefore cannot answer how hippocampus and entorhinal cortex differentially contribute to context inference, how the sparsity of Purkinje cell activity affects it, or what loss of specific cell types would do.