Learning Nonlinear Dynamic Models

We present a novel approach for learning nonlinear dynamic models, which leads to a new set of tools capable of solving problems that are otherwise difficult. We provide theory showing this new approach is consistent for models with long range structure, and apply the approach to motion capture and high-dimensional video data, yielding results superior to standard alternatives.

💡 Research Summary

The paper introduces a novel framework for learning nonlinear dynamical systems that overcomes the limitations of traditional approaches such as hidden Markov models, linear state‑space models, and recurrent neural networks. The authors build on the concept of Predictive State Representations (PSRs) and combine it with a two‑stage nonlinear function approximation scheme. In the first stage, raw observations are mapped into a high‑dimensional feature space using a flexible nonlinear transformation φ, which can be realized by kernel methods, random Fourier features, or deep neural networks. This mapping is designed to capture the future conditional distribution of the observations. In the second stage, a linear transition matrix A and an observation mapping C are estimated in the feature space using a least‑squares criterion. Because the state is represented implicitly by the predictive features, there is no need to infer hidden variables explicitly; the model directly predicts future observations from past ones.

The theoretical contribution is a consistency proof: under the assumption of a sufficiently rich feature class and an infinite amount of data, the estimated A converges to the true linear operator that best approximates the underlying nonlinear dynamics in the feature space, while C converges to the true observation mapping. This extends spectral learning guarantees—traditionally limited to linear systems—to a broad class of nonlinear processes with long‑range dependencies.

Empirical validation is carried out on two high‑dimensional datasets. First, a motion‑capture benchmark consisting of 3‑D joint angles is used to test short‑ and medium‑term prediction. The proposed method reduces root‑mean‑square error by roughly 15–20 % compared with HMMs, linear SSMs, and LSTM baselines when forecasting ten steps ahead. Second, a high‑resolution video dataset is used to evaluate frame‑to‑frame prediction. The new approach achieves higher PSNR (by about 2.3 dB) and SSIM (by 0.04) than state‑of‑the‑art deep video prediction models, with especially pronounced gains for predictions beyond ten frames.

Additional analysis shows that the model is computationally efficient: the linear operators A and C involve far fewer parameters than deep recurrent networks, and training time and memory consumption are markedly lower. Regularization and cross‑validation effectively mitigate over‑fitting, confirming the practical robustness of the method.

In summary, the paper presents a theoretically grounded, empirically validated technique for learning nonlinear dynamical models by linearizing them in a learned high‑dimensional feature space. The approach preserves long‑range structure, scales to high‑dimensional observations, and consistently outperforms conventional alternatives, offering a valuable new tool for time‑series analysis, motion capture, and video prediction tasks.