The minimal computational substrate of fluid intelligence.

Area of Science:

• Cognitive science
• Artificial intelligence
• Neuroscience

Background:

• Quantifying cognitive abilities relies on behavioral tasks, but task specificity and generalizability are often uncertain.
• Raven's Advanced Progressive Matrices (RAPM) is a common clinical test for fluid intelligence.

Purpose of the Study:

• To evaluate if a self-supervised artificial neural network (LaMa) trained on natural scenes can solve RAPM tests.
• To investigate if LaMa's performance and error patterns mimic human performance, particularly in relation to brain function.

Main Methods:

• LaMa, a self-supervised neural network, was trained on image completion tasks.
• LaMa's performance on a compact RAPM was compared to human participants (healthy and brain-lesioned).
• Error patterns of LaMa were analyzed under conditions of degraded spatial pattern integration.

Main Results:

• LaMa achieved human-level scores on the RAPM test without task-specific training or inductive bias.
• LaMa demonstrated human-like item difficulty variation and error patterns.
• LaMa's errors under spatial integration challenges resembled those seen with right frontal lobe damage.

Conclusions:

• Matrix-style intelligence tests may be susceptible to computationally simple solutions.
• These solutions may not necessarily engage the neural substrates typically associated with reasoning.
• LaMa's performance challenges assumptions about the cognitive processes underlying fluid intelligence tests.