Speaker
Description
Reservoir computing (RC) provides an efficient framework for data-driven modeling of nonlinear dynamics, yielding an autonomous dynamical system that faithfully reproduces the driving dynamics. However, where contextual labels are not present, standard RC learns an independent readout for each time series, with no shared structure across tasks, no generalization to unseen parameters, and no sample efficiency across families of related systems. To overcome these limitations, we developed a modal decomposition describing each system's readout as a linear combination of a shared basis in the readout space, with task-specific mixing coefficients acting as latent variables. Both basis and coefficients are inferred jointly from raw data via expectation maximization with closed-form steps. The reservoir serves as a canonicalizing map: heterogeneous input sequences are projected into a common feature space where cross-task variability reduces to a low-dimensional structure. After orthonormalization, each basis element captures an independent mode of variation, and the latent coordinates spontaneously organize according to the governing physical parameters, enabling interpolation and extrapolation to unseen dynamical regimes in closed-loop generation. We first validate the methodology on families of dynamical systems and on a public dataset of locomotion patterns. We then apply this framework to build patient-specific digital twins of sensorimotor behavior in a clinical cohort of healthy, mildly cognitively impaired, and Alzheimer's disease participants performing reaching and catching movements in virtual reality. Each trial is compressed into a handful of coefficients that faithfully reproduce the recorded kinematics in closed loop. At the trial level, the coefficients encode kinematic features of the movement. Crucially, the latent-space organization degrades with cognitive impairment: the geometry of the distribution of the individual coefficients progressively deteriorates, reflecting reduced motor planning precision. Learned entirely without clinical supervision, these signatures separate diagnostic groups and track cognitive scores, yielding an interpretable dynamical biomarker of neurodegeneration derived from motor behavior.
| Preferred Presentation | Oral Presentation |
|---|