Human-like dissociations between confidence and accuracy in convolutional neural networks

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Artificial Intelligence

Background:

• Human confidence-accuracy dissociations occur when stimulus energy (contrast, variability) changes, with higher energy leading to higher confidence even with matched performance.
• The positive evidence heuristic is a common explanation, suggesting confidence ignores disconfirming evidence.
• The signal-and-variance-increase hypothesis offers an alternative, attributing dissociations to changes in perceptual representation separation and variance.

Purpose of the Study:

• To test if confidence-accuracy dissociations emerge naturally in artificial neural networks (ANNs), specifically convolutional neural networks (CNNs), which lack built-in confidence heuristics.
• To determine if ANNs can replicate human confidence-accuracy dissociations induced by stimulus energy manipulations.
• To investigate the underlying mechanisms of these dissociations in CNNs and compare them to human cognitive models.

Main Methods:

• Convolutional neural networks (CNNs) of varying architectures (shallow to deep, including VGG-19 and ResNet-50) were tested.
• Stimulus energy was manipulated using contrast and variability.
• Confidence-accuracy relationships were analyzed across different CNNs and stimulus conditions.

Main Results:

• CNNs consistently produced confidence-accuracy dissociations mirroring human findings across various stimulus energy manipulations.
• These dissociations were observed across a range of CNN architectures, from simple to complex.
• The mechanism identified in CNNs was an increase in signal separation and variance in the output layer, consistent with the signal-and-variance-increase hypothesis.

Conclusions:

• Confidence-accuracy dissociations can arise in artificial systems without explicit confidence heuristics, challenging the necessity of the positive evidence heuristic in humans.
• The signal-and-variance-increase mechanism appears to be a fundamental driver of confidence-accuracy dissociations, applicable to both artificial and biological systems.
• CNNs serve as valuable models for testing and refining cognitive theories of human perception and confidence.