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

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Artificial Intelligence

Background:

• Human confidence-accuracy relationships are often dissociated by stimulus energy (contrast, variability).
• The positive evidence heuristic is a common explanation, but the signal-and-variance-increase hypothesis offers an alternative.
• Artificial neural networks (ANNs) provide a model system to test the necessity of cognitive heuristics.

Approach:

• Investigated confidence-accuracy dissociations in convolutional neural networks (CNNs) under stimulus energy manipulations.
• Tested various CNN architectures, from shallow to deep networks (e.g., VGG-19, ResNet-50) pretrained on ImageNet.
• Analyzed internal representations to identify the source of confidence-accuracy dissociations.

Key Points:

• CNNs naturally produced confidence-accuracy dissociations mirroring human behavior across different energy manipulations.
• These dissociations were consistent across diverse CNN architectures.
• The underlying cause in CNNs was identified as increased separation and variance in internal representations due to higher stimulus energy.

Conclusions:

• Confidence-accuracy dissociations can arise in ANNs without explicit confidence heuristics, suggesting they are not strictly necessary for this phenomenon.
• The findings support the signal-and-variance-increase hypothesis as a potential explanation for confidence-accuracy dissociations in both humans and artificial systems.
• CNNs serve as valuable models for distinguishing between low-level, stimulus-driven and high-level, cognitive explanations of behavior.