Using artificial neural networks to reveal the human confidence computation

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Artificial Intelligence

Background:

• Human confidence judgments are crucial for decision accuracy but poorly understood for naturalistic, multi-choice tasks.
• Existing cognitive models are limited, primarily addressing simple, two-choice scenarios.

Purpose of the Study:

• To investigate mechanisms of confidence in multi-alternative, naturalistic decision-making.
• To compare various confidence strategies using a novel convolutional neural network (CNN) model.

Main Methods:

• Utilized a CNN model (RTNet) to analyze confidence strategies in an eight-alternative digit discrimination task (N=60).
• Compared strategies based on evidence distribution (full vs. subset) and evidence type (posterior vs. raw).

Main Results:

• The Top2Diff model, using raw evidence difference between the top two choices, showed superior quantitative and qualitative fits.
• This model also provided the best predictions of human confidence ratings.

Conclusions:

• Human confidence appears to rely on a specific subset of evidence, challenging existing theories like the Bayesian confidence hypothesis.
• CNNs offer a powerful framework for modeling human confidence in complex decision-making environments.