Relational reasoning and generalization using nonsymbolic neural networks

Area of Science:

• Cognitive Science
• Artificial Intelligence
• Neuroscience

Background:

• Abstract relational reasoning is crucial for cognition.
• Previous research indicated neural networks struggle with representing mathematical identity (equality).
• This limitation suggested a gap between artificial and human abstract reasoning.

Purpose of the Study:

• To investigate neural networks' capacity for learning and generalizing equality (identity).
• To determine if neural networks can perform tasks previously considered uniquely human symbolic abilities.
• To explore the emergence of symbolic reasoning from data-driven learning.

Main Methods:

• Assessed out-of-sample generalization of equality learning in neural networks.
• Utilized arbitrary and pre-trained representations for equality tasks.
• Tested models on basic, sequential (ABA-pattern), and hierarchical equality problems.

Main Results:

• Neural networks successfully learned basic mathematical identity.
• Models demonstrated generalization on sequential and complex hierarchical equality tasks, even with limited training data (zero-shot).
• Performance on these tasks matched or exceeded benchmarks for human-unique symbolic abilities.

Conclusions:

• Neural networks are capable of learning and generalizing abstract equality, contrary to prior beliefs.
• Symbolic reasoning capabilities can emerge through data-driven, non-symbolic learning processes in neural networks.
• These findings bridge the gap between connectionist and symbolic approaches to artificial intelligence and cognitive science.