A Semi-parametric Model for Decision Making in High-Dimensional Sensory Discrimination Tasks

Stephen Keeley, Benjamin Letham, Craig Sanders, Chase Tymms, Michael ShvartsmanAAAI 2023 · OpenReview

Abstract

Psychometric functions typically characterize binary sensory decisions along a single stimulus dimension. However, real-life sensory tasks vary along a greater variety of dimensions (e.g. color, contrast and luminance for visual stimuli). Approaches to characterizing high-dimensional sensory spaces either require strong parametric assumptions about these additional contextual dimensions, or fail to leverage known properties of classical psychometric curves. We overcome both limitations by introducing a semi-parametric model of sensory discrimination that applies traditional psychophysical models along a stimulus intensity dimension, but puts Gaussian process (GP) priors on the parameters of these models with respect to the remaining dimensions. By combining the flexibility of the GP with the deep literature on parametric psychophysics, our semi-parametric models achieve good performance with much less data than baselines on both synthetic and real-world, high-dimensional psychophysics datasets. We additionally show strong performance in a Bayesian active learning setting, and present a novel active learning paradigm for the semi-parametric model.

Introduction

In psychophysics, researchers aim to understand the relationship between physical stimuli and mental perception. This is traditionally modeled using a Psychometric Function (PF), a sigmoidal curve representing the probability of detection as a function of stimulus intensity. However, real-world stimuli are inherently high-dimensional, involving attributes like pitch, timbre, color, and contrast.

Existing approaches to high-dimensional sensory mapping face a trade-off:

  1. Parametric models are often too rigid, requiring a priori knowledge of the mathematical interactions between dimensions.
  2. Non-parametric models, such as standard Gaussian Processes (GPs), are flexible but lack the "physical" constraints of sensory perception, such as the monotonic relationship between intensity and detection probability.

Methodology

The proposed Semi-parametric Psychophysical Model addresses these issues by partitioning the stimulus space into intensity ($i$) and context ($s$). The intensity dimension is governed by a traditional sigmoidal link function (e.g., Probit or Logit). Crucially, the parameters of this sigmoid—specifically the slope ($k$) and offset ($c$)—are not fixed but are treated as functions of the contextual dimensions.

Graphical model of the semi-parametric psychophysical model.

By placing Gaussian Process (GP) priors on these parameters, the model creates a "psychometric field" that adapts smoothly across the contextual space. To handle the resulting mathematical intractability, the authors developed two inference strategies: Factorized Variational Inference for robust learning and an MVN Approximation for faster computation and compatibility with active learning tools.

Experimental Setup

The model's performance was evaluated across several datasets:

  • Synthetic Tasks: 2D and 8D simulated environments used to benchmark learning speed and accuracy against ground truth.
  • Human 2AFC Dataset: 3,000 trials of a two-alternative forced choice visual task involving contrast, luminance, and frequency.
  • Human 4AFC Dataset: Data from four subjects performing a four-alternative forced choice task (200–500 trials) with Gabor patches varying in size, orientation, and color.

Results and Discussion

A key finding is that the inductive bias provided by the sigmoidal structure along the intensity dimension significantly improves performance. While a "pure" GP must learn the concept of a detection threshold from scratch, the semi-parametric model leverages this known psychophysical property.

Comparison of response probability curves across different models.

As shown in the figure above, the semi-parametric model produces curves that are both flexible and physically plausible, unlike the unconstrained GP-RBF or the overly restrictive GP-linear models.

Key Performance Metrics

  • Sample Efficiency: In 8D synthetic tasks, the semi-parametric model achieved high accuracy using approximately half the samples required by standard GP baselines.
  • Real-World Accuracy: On human behavioral data, the model consistently outperformed both linear and unconstrained GP models in cross-validated log-likelihood across all subjects.
  • Active Learning: The model's explicit parameterization of the detection threshold enabled a new strategy called ThresholdBALV, which identifies the most informative stimuli for human participants, significantly accelerating experimental mapping.

Conclusion

The semi-parametric approach bridges the gap between rigid parametric models and overly flexible non-parametric methods. By incorporating known psychophysical constraints into a Gaussian Process framework, it offers a powerful tool for mapping high-dimensional sensory spaces with high sample efficiency. Future work may explore hierarchical versions of the model to share information across subjects and extend the framework to categorical tasks without monotonic intensity dimensions. The authors have integrated this model into the AEPsych library for community use.

Summary updated

Abstract

Psychometric functions typically characterize binary sensory decisions along a single stimulus dimension. However, real-life sensory tasks vary along a greater variety of dimensions (e.g. color, contrast and luminance for visual stimuli). Approaches to characterizing high-dimensional sensory spaces either require strong parametric assumptions about these additional contextual dimensions, or fail to leverage known properties of classical psychometric curves. We overcome both limitations by introducing a semi-parametric model of sensory discrimination that applies traditional psychophysical models along a stimulus intensity dimension, but puts Gaussian process (GP) priors on the parameters of these models with respect to the remaining dimensions. By combining the flexibility of the GP with the deep literature on parametric psychophysics, our semi-parametric models achieve good performance with much less data than baselines on both synthetic and real-world, high-dimensional psychophysics datasets. We additionally show strong performance in a Bayesian active learning setting, and present a novel active learning paradigm for the semi-parametric model.
BibTeX
@inproceedings{keeley2023semi,
  author = {Stephen Keeley and Benjamin Letham and Craig Sanders and Chase Tymms and Michael Shvartsman},
  title = {{A Semi-parametric Model for Decision Making in High-Dimensional Sensory Discrimination Tasks}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2023},
  url = {https://openreview.net/forum?id=DK3NSMBdEJg},
}

Loading related content...

Figure