Cerebellar inspiration, contemplative neuroscience, and a benchmark for event vision

The cerebellum has long been the brain region that neuroscience research engineers gesture toward when asked which biological structure looks most “designed for control.” Its granular layer expands inputs into a high-dimensional sparse representation; its Purkinje cells integrate this expansion under climbing-fiber supervision; the resulting circuit is implicated in motor learning, predictive coding, and increasingly, the kind of internal-model learning that RL needs.

CDRL operationalizes the circuit literally. The actor is a granule-layer-style sparse high-dimensional projection; the critic is a Purkinje-style supervised integrator that receives an error signal styled on climbing-fiber dynamics. The training scheme keeps the analogy intact: errors are sparse and localized rather than backpropagated densely.

On a battery of MuJoCo continuous-control tasks, CDRL reaches PPO-equivalent performance with 3.4× fewer environment steps. The architecture is small — the granule projection is the only large layer — which suggests the gain is structural, not parametric. The authors are careful to flag that the cerebellar analogy is loose at the cellular level; the win is at the level of the circuit topology and the locality of the error.

This is a contemplative-neuroscience paper that the AI audience should pay attention to, for an unusual reason. Jhana states — deep absorption states from the Theravada Buddhist contemplative tradition — have historically been difficult to study because they are rare, are reported only by experienced practitioners, and their neural signatures (if any) have been speculative.

The work pools data from 47 advanced practitioners across three meditation centers, totaling 312 sessions with synchronized EEG. A relatively small ensemble model — a few dense layers fed by power-spectral features — reaches 89% classification accuracy distinguishing jhanic from non-jhanic meditative states, against a chance baseline of 50%. The signature the model relies on is a specific gamma-band coherence pattern across parietal channels that does not appear in baseline attention.

The relevance for AI interpretability is indirect but interesting: it offers a model organism for the question of whether unusual internal states (states the practitioner reports but cannot easily verify) leave robust external signatures (signal patterns that hold up across individuals and sessions). That is structurally the same question facing interpretability research on language models, which is increasingly being asked to characterize internal model states that resist direct probing.

Event-based vision has suffered from a benchmark problem. The existing datasets — DVS128, N-MNIST, N-Caltech — were designed to demonstrate that event cameras could be used for classification at all. They are too small and too geometrically simple to differentiate methods that actually exploit event structure from methods that merely re-render events to frames and apply standard ConvNets.

EPRBench takes the opposite stance. It is structured around tasks that punish the re-render-to-frames shortcut: temporal-order reasoning, ego-motion-conditional object permanence, and prediction of object trajectories from sparse event subsets that lack the frame-equivalent’s visual completeness. Baseline numbers are calibrated to make the shortcut visible: a strong frame-rendering ConvNet sits ~22 points below a true event-stream-native model.

This is the kind of benchmark that the neuromorphic vision community has been asking for and has so far had to defer. The hard version of the question — what does an event-driven system buy you that a frame-driven one fundamentally cannot — finally has a place to be argued empirically.