Physiological Signal Imputation using Reservoir Computing Energy-Based Models

NLDL 2026 Conference Submission70 Authors

Keywords: Reservoir Computing, Energy Based Models, Principle Component Analysis, Uncertainty Estimation, Liquid State Machine, Echo State Network, Manifold

TL;DR: We propose a novel architecture that combines reservoir computing with an energy-based model, evaluating which features best support the EBM by comparing reservoirs parameterized by neuron models and their lower-dimensional manifolds.

Abstract: Many biological time-series datasets (e.g., neural recordings and physiological signals) have missing segments due to sensor dropouts or noise, which leads to a need for a robust imputation method that also quantifies confidence. We propose an architecture using reservoir computing (Echo State Networks and Liquid State Machines) in series with an Energy-Based Model (EBM) to perform the imputation. An EBM models the underlying data distribution by assigning a scalar energy to possible outputs, with a low energy for plausible imputations and high energy for implausible ones. Therefore, the energy serves as a heuristic for uncertainty of the prediction, meaning unreliable imputations may be detected by their higher energy. To capture the complex temporal dynamics of biological data, we embed inputs into a high-dimensional state space using reservoir computing, either using Echo State Networks (ESNs) or Liquid State Machines (LSMs). ESNs have fixed random recurrent weights, and LSMs have spiking neurons. Both these reservoirs expand and filter the input sequences into rich nonlinear features, simplifying the imputation task for the EBM. The output of the reservoir is passed into the EBM using two different methods. Either raw signals from the reservoir are used as inputs or we perform a dimensionality reduction of the reservoir states, via principal component analysis (PCA), to identify low-dimensional manifolds of temporal patterns, which can denoise the reservoir representation to create a cleaner signal for the EBM. Preliminary results on ECG signals show that using an LSM as the reservoir or the lower dimensional manifold of an ESN produces the best features for the EBM to do signal imputation.

Serve As Reviewer: ~Robert_Clarke1

Submission Number: 70