Abstract
Traditional linear models often struggle to capture "temporal diffusion"—the lingering influence of past stimuli on current responses. We present Deconvolutional Time Series Regression (DTSR), a framework that models these effects by learning latent Impulse Response Functions (IRFs). Evaluated on psycholinguistic datasets, DTSR demonstrates superior predictive accuracy and provides interpretable insights into cognitive dynamics, such as the distinction between perceptual and high-level cognitive processing speeds.
Introduction
In fields like psycholinguistics, researchers study how stimuli (e.g., reading a word) affect behavioral responses (e.g., fixation duration). However, human cognition is characterized by "spillover" effects, where the difficulty of a previous stimulus affects current performance. Standard approaches using lagged regressors are often arbitrary and computationally expensive. DTSR addresses these flaws by treating predictors as a continuous stream of impulses convolved with learned decay functions.
Methodology
DTSR utilizes the concept of convolution from signal processing. Instead of static predictors, the model learns an Impulse Response Function (IRF) for each covariate, typically using a ShiftedGamma kernel.
The response at any given time is the sum of these convolved signals. Implementation via TensorFlow and Edward allows for automatic differentiation and Black Box Variational Inference (BBVI), enabling the model to handle complex IRF shapes and random effects efficiently.
Results and Discussion
DTSR was tested on synthetic data and three human corpora (Natural Stories, Dundee, and UCL). It significantly outperformed Linear Mixed Effects (LME) models in predictive power ($p = 0.0001$).
Cognitive Signatures
The model revealed a "cognitive asymmetry": low-level variables like word length decay quickly, whereas high-level variables like "surprisal" have long-lasting effects. Furthermore, the model identified a "Rate" effect, suggesting motor inertia in self-paced reading tasks.
Cross-Corpus Consistency
A key finding was the stability of the learned IRFs across different datasets. As shown below, the IRFs for the Dundee and UCL eye-tracking corpora are remarkably consistent, suggesting that DTSR captures fundamental, replicable cognitive signatures.
Limitations
While powerful, DTSR relies on non-convex stochastic optimization, which can lead to local optima. Consequently, derived p-values should be considered approximate. Additionally, convolving long histories is more computationally intensive than standard regression.
Conclusion
DTSR provides a robust, parsimonious method for modeling temporally diffuse effects. Its ability to recover latent IRFs offers a new window into the timing of cognitive processes, with potential applications extending to fMRI and EEG analysis.
Summary updated