[Paper Review] Perceptual decision making: Biases in post-error reaction times explained by attractor network dynamics

Kevin Berlemont, Jean‐Pierre Nadal|arXiv (Cornell University)|Mar 2, 2018

Neural dynamics and brain function93 references22 citations

TL;DR

This paper proposes that post-error slowing and post-error improvement in accuracy—behavioral phenomena observed in perceptual decision-making without feedback—arise from intrinsic nonlinear dynamics in a reduced attractor network model. By incorporating a post-decision inhibitory current, the model reproduces these sequential effects qualitatively and with correct magnitude, demonstrating that they emerge from network dynamics rather than external feedback or explicit memory mechanisms.

ABSTRACT

Perceptual decision making is the subject of many experimental and theoretical studies. Whereas most modeling analysis are based on statistical processes of accumulation of evidence, less attention is being devoted to the modeling with attractor network dynamics, even though they describe well psychophysical and neurophysiological data. In particular, very few works confront attractor models predictions with data from continuous sequences of trials. Recently however, a biophysical competitive attractor network model has been used to describe such sequences of decision trials, and has been shown to reproduce repetition biases observed in perceptual decision experiments. Here we propose an extension of the reduced attractor network model of Wong and Wang (2006) to get more insights into such effects. We make explicit the conditions under which such network can perform a succession of decisions, and show that the model provides a mathematical framework for studying the effects of a trial on the decision made on the next one. We study in details the reaction times properties during a sequence of decision trials, and show that the model reproduces behavioral data, both qualitatively and quantitatively. In particular, we find that the decision made on the current trial is biased toward the one made on the previous trial. More remarkably, we show that, in the absence of any feedback about the correctness of the decision, the network exhibits post-error slowing, a subtle effect in agreement with empirical data.

Motivation & Objective

• To explain post-error slowing and post-error improvement in accuracy in perceptual decision-making without relying on feedback or explicit memory units.
• To investigate how intrinsic nonlinear dynamics in attractor networks give rise to sequential effects in continuous decision tasks.
• To determine the conditions under which an inhibitory input after decision-making enables stable, repeated trial performance without trapping in attractors or losing memory of past states.
• To compare model predictions with empirical data on reaction times and accuracy in sequential binary choice tasks.
• To demonstrate that first-order sequential effects (e.g., post-error adjustments) emerge naturally from network dynamics, even without additional memory modules.

Proposed method

• Adaptation of the reduced Wong and Wang (2006) attractor network model to include a post-decision inhibitory current.
• Incorporation of a biophysically motivated inhibitory input triggered after each decision to reset network activity.
• Mathematical analysis of the network’s nonlinear dynamics to identify parameter regimes enabling stable trial succession.
• Numerical simulations of continuous sequences of two-alternative forced-choice (TAFC) trials with inter-trial intervals as short as 500 ms.
• Analysis of reaction times and error rates across sequences to assess post-error and post-correct dynamics.
• Comparison of model outputs with empirical behavioral data on post-error slowing and post-error accuracy improvement.

Experimental results

Research questions

• RQ1 Can a minimal attractor network model reproduce post-error slowing and post-error improvement in accuracy without feedback or explicit memory units?
• RQ2 What are the dynamical conditions under which an inhibitory input after decision-making enables stable, repeated trial performance without attractor trapping or memory loss?
• RQ3 How do reaction times and accuracy depend on the nonlinear dynamics of the network during sequential decision-making?
• RQ4 To what extent do first-order sequential effects emerge as intrinsic properties of the network dynamics, rather than from additional mechanisms?
• RQ5 Why do post-error effects persist even when subjects are unaware of their errors, based on the network’s dynamic states?

Key findings

• The model reproduces post-error slowing (longer reaction times after errors) and post-error improvement in accuracy (reduced error rates after errors) with qualitatively correct behavior and correct orders of magnitude.
• These effects emerge naturally from the nonlinear dynamics of the attractor network with a post-decision inhibitory input, without requiring feedback or explicit memory modules.
• The network avoids being trapped in the first attractor or losing memory of past dynamics when the inhibitory input is appropriately tuned.
• For parameters yielding first-order sequential effects (e.g., ICD,max = 0.035 nA), higher-order effects such as second-order biases or complex repetition patterns are not reproduced, indicating a limit of the minimal model.
• Post-error and post-correct firing rate distributions overlap significantly, making single-trial error detection difficult, which may explain why errors are often not consciously perceived.
• The model suggests that post-error effects are not due to strategic adjustments but are instead inherent to the network’s relaxation dynamics following decision states.