Human-like systematic generalization through a meta-learning neural network.

Area of Science:

• Cognitive Science
• Artificial Intelligence
• Computational Linguistics

Background:

• Human cognition relies on systematic compositionality, enabling novel combinations of known elements.
• Fodor and Pylyshyn's challenge posits that artificial neural networks lack this systematicity, limiting their viability as models of the mind.
• Despite advancements, achieving systematic generalization in neural networks remains a persistent challenge.

Purpose of the Study:

• To demonstrate that neural networks can achieve human-like systematicity.
• To introduce and evaluate the meta-learning for compositionality (MLC) approach.
• To compare the generalization capabilities of MLC with other models and human performance.

Main Methods:

• Developed the meta-learning for compositionality (MLC) approach, guiding neural network training with diverse compositional tasks.
• Conducted human behavioral experiments using an instruction learning paradigm.
• Evaluated seven different models, including MLC, probabilistic symbolic models, and standard neural networks, on systematic generalization benchmarks.

Main Results:

• MLC successfully achieved both systematicity and flexibility, outperforming rigid symbolic models and unsystematic neural networks.
• MLC demonstrated human-like generalization capabilities in head-to-head comparisons.
• MLC significantly advanced the compositional skills of machine learning systems across several benchmarks.

Conclusions:

• Optimizing neural networks for compositional skills enables human-like systematic generalization.
• The MLC approach provides a viable method for developing more capable and human-like artificial intelligence.
• This research bridges the gap between artificial neural networks and the systematic nature of human thought and language.