Compositionality is the demand that intelligence preserve structure: complex thoughts or linguistic expressions should be understood through discrete parts and the rules that combine them. In the classical symbolic view, reasoning is specified explicitly in logic, making compositional structure clear but potentially brittle when real-world details or grounding are underspecified. section 7.6
Connectionists challenge the assumption that compositionality must be programmed directly. They learn statistical regularities through gradient-based methods and distributed representations; critics such as Fodor and Pylyshyn argue that this can miss the structure of thought and language. Smolensky’s reply is that symbolic-like rules may emerge as higher-level descriptions of neural systems governed by softer constraints. section 7.6
A testable disagreement
The dispute is not settled by philosophical preference alone. Standardized benchmarks, shared datasets, and measured improvements over strong baselines test whether learned systems achieve useful reasoning without prespecified symbolic programs. section 7.6
