Understanding the structure of cognitive noise

Jian-Qiao Zhu, Samuel J. Gershman, Pablo León-Villagrá, Nick Chater, Adam N. SanbornPLOS Computational Biology 2022 · OpenReview

TL;DR

This paper reveals that "cognitive noise" – the seemingly random fluctuations in our judgments and responses – isn't random at all. It exhibits a surprising structure, characterized by both heavy-tailed distributions and long-range autocorrelations (1/f noise), which co-occur across different cognitive tasks. This structured noise might be a functional adaptation for navigating complex environments, as demonstrated by a statistical sampling algorithm that mimics these properties.

What Problem Does This Paper Address?

In cognitive science, "noise" is typically treated as an undesirable random error, a nuisance to be averaged out in experiments. Standard models often assume this noise follows a simple Gaussian distribution, implying independence between successive responses. However, previous research has hinted at a more complex reality:

  1. Heavy-tailed distributions: Response times in tasks like recalling animal names show that while most intervals between responses are short, there are occasional very long ones. This pattern is reminiscent of animal foraging strategies in patchy environments.
  2. 1/f noise (long-range autocorrelations): In tasks like estimating durations, responses aren't just influenced by the immediately preceding one but also by a much longer history of past responses. This suggests a more integrated, less independent process.

The problem is that these two findings, heavy tails and 1/f noise, have largely been studied in isolation. Existing models that explain one often fail to predict the other, leaving a gap in our understanding of the overall structure of cognitive noise and its potential functional role.

Methods

The researchers employed two main experimental paradigms and a computational modeling approach:

  • Experimental Tasks:
    • Animal Naming Task: Participants were asked to list animal names as they came to mind, with the constraint of not repeating names consecutively. This task was previously used to demonstrate heavy-tailed inter-response intervals (IRIs).
  • Time Estimation Task: Participants were asked to reproduce specific time intervals (e.g., 1/3s, 1s, 3s) by repeatedly pressing a key. This task was previously used to demonstrate 1/f noise in response times.
  • Data Analysis:
    • Heavy Tails: They fitted power-law distributions to the inter-response intervals (IRIs) and time estimates to quantify the tail exponent ($\mu$).
  • 1/f Noise: They analyzed the power spectral density of the response time sequences to estimate the autocorrelation exponent ($\alpha$), looking for the characteristic 1/f scaling.
  • Computational Modeling:
    • They investigated three Markov Chain Monte Carlo (MCMC) sampling algorithms: Random Walk Metropolis (RWM), Hamiltonian Monte Carlo (HMC), and Metropolis-coupled MCMC (MC³). These algorithms are used in statistics to approximate solutions to complex probabilistic problems by drawing samples from a distribution.
  • They modeled the cognitive tasks as sampling processes in a "hypothesis space" (e.g., semantic space for animal names, time estimates for duration).
  • They used Approximate Bayesian Computation (ABC) to compare how well each MCMC algorithm's simulated output matched the observed human data (both heavy tails and 1/f noise).

Datasets

The study utilized two datasets collected from human participants:

  1. Animal Naming Dataset: Data from 10 participants performing the animal naming task.
  2. Time Estimation Dataset: Data from 30 participants performing the time estimation task across three different target time intervals.

The authors also generated synthetic data from the MCMC sampling algorithms to compare against the human datasets.

Key Insights and Ideas

  • Co-occurrence of Noise Features: The most striking finding is that heavy-tailed distributions and 1/f noise are not mutually exclusive in human cognition. They appear together in tasks where they were previously thought to be distinct. This challenges existing models that treat these phenomena separately.
  • Noise as a Functional Adaptation: The paper proposes that this structured noise might be a functional adaptation for dealing with complex, "patchy" environments (like semantic memory or temporal uncertainty). The ability to have occasional large deviations (heavy tails) while maintaining long-term dependencies (1/f noise) could be adaptive.
  • MCMC as a Cognitive Model: The study suggests that MCMC sampling algorithms, particularly MC³, can serve as plausible computational models for cognitive noise. This is significant because MCMC algorithms are designed to handle complex, potentially multimodal probability distributions, which aligns with the idea of navigating "patches" of information.
  • "Intuitive Statistician" Revisited: The findings lend support to the idea that the brain acts as an "intuitive statistician," using approximation methods (like sampling) to solve complex probabilistic problems, rather than relying on perfect, computationally intractable Bayesian inference.

Key Results

  • Human Data:
    • In the animal naming experiment, participants showed both heavy-tailed IRIs ($\mu$ in [1.29, 2.61]) and 1/f noise ($\alpha$ in [0.38, 0.90] for 9 out of 10 participants).
  • In the time estimation experiment, participants also exhibited both 1/f noise ($\alpha$ values often in the [0.53, 2.06] range) and heavy-tailed trial-by-trial changes ($\mu$ in [0.87, 2.67] for 23 out of 30 participants).
  • Crucially, a significant number of individual participants showed evidence of both heavy tails and 1/f noise in their data.
  • Model Performance:
    • The MC³ algorithm provided the best fit to the human data in the time estimation task, significantly outperforming RWM and HMC.
  • For the animal naming task, RWM and MC³ performed comparably, with RWM fitting slightly more participants.
  • When comparing the model's simulated output to human data, MC³ showed a strong correspondence in capturing both heavy tails and 1/f noise, particularly in the animal naming task.

Limitations and Future Work

  • Model Complexity: While MC³ is a good fit, the paper acknowledges that it's a complex algorithm. The question remains whether simpler sampling strategies could also explain the observed phenomena.
  • Unimodality Assumption: The study notes that while MC³ was designed for multimodal distributions, it also performed well on the time estimation task, which was assumed to be unimodal. This suggests that the brain might employ such algorithms even when not strictly necessary, perhaps due to the difficulty in definitively identifying unimodality.
  • Functional Role: While the paper suggests noise is a functional adaptation, further research is needed to explore the precise benefits of this structured noise in different cognitive domains and under varying environmental conditions.
  • Alternative Explanations: The authors mention that other accounts, like Brownian motion in multifractal time, could potentially describe the co-occurrence of heavy tails and 1/f noise, but these models may not fully explain task performance (e.g., accurate time estimation). Future work could explore these alternative theoretical frameworks.
  • Generalizability: The study used two specific tasks. Investigating whether this co-occurrence of noise features and the MC³ model hold for a wider range of cognitive tasks would be valuable.

Summary updated

Abstract

Human cognition is fundamentally noisy. While routinely regarded as a nuisance in experimental investigation, the few studies investigating properties of cognitive noise have found surprising structure. A first line of research has shown that inter-response-time distributions are heavy-tailed. That is, response times between subsequent trials usually change only a small amount, but with occasional large changes. A second, separate, line of research has found that participants’ estimates and response times both exhibit long-range autocorrelations (i.e., 1/f noise). Thus, each judgment and response time not only depends on its immediate predecessor but also on many previous responses. These two lines of research use different tasks and have distinct theoretical explanations: models that account for heavy-tailed response times do not predict 1/f autocorrelations and vice versa. Here, we find that 1/f noise and heavy-tailed response distributions co-occur in both types of tasks. We also show that a statistical sampling algorithm, developed to deal with patchy environments, generates both heavy-tailed distributions and 1/f noise, suggesting that cognitive noise may be a functional adaptation to dealing with a complex world.
BibTeX
@article{zhu2022understanding,
  author = {Jian-Qiao Zhu and Samuel J. Gershman and Pablo León-Villagrá and Nick Chater and Adam N. Sanborn},
  title = {{Understanding the structure of cognitive noise}},
  journal = {PLOS Computational Biology},
  year = {2022},
  url = {https://openreview.net/forum?id=QP90Vy8dgd},
}
Figure