Physiological basis of performance degradation in living neural reservoirs

Determine the physiological mechanisms responsible for the post‑training degradation of classification accuracy in reservoir computing experiments using optogenetically stimulated motoneuron cultures on microelectrode arrays, specifically the drift in network dynamics that collapses initially separable latent trajectories into poorly separable clusters within a few hours after training.

Background

The paper observes that living neural cultures (particularly Types C/D) initially achieve high pattern‑classification accuracy with reservoir computing but experience a marked decline after 1–2 hours, with substantial degradation by ~4 hours. Latent trajectory analysis shows that stimulus‑evoked neural trajectories, initially well separated, collapse into overlapping clusters, reducing decodability.

The authors suggest possible biological drivers such as synaptic fatigue, plasticity, ion‑channel adaptation, and/or excitotoxicity, but explicitly state that the physiological basis for the observed degradation remains unclear. Identifying the mechanisms would inform strategies to maintain stability and prolong reliable computational performance beyond the current time window.

References

The physiological basis of this degradation remains unclear and likely involves a range of factors (as discussed below and in SI~\ref{si:failure-modes}).

— Computing with Living Neurons: Chaos-Controlled Reservoir Computing with Knowledge Transplant  (2604.02552 - Kim et al., 2 Apr 2026) in Main text — Subsection "Reconstruction task with (naive) Reservoir Computing", paragraph preceding Section "Chaos-controlled Reservoir Computing (cc-RC)" (discussion around Fig. 3d)