In the standard Hopfield model, stored patterns are random — each bit is independent. Real patterns have internal structure: correlations, symmetries, subgroups. A face has two eyes. A word has phonetic regularities. The pattern is not just its bits but the relationships between them.
The paper studies what happens when structured patterns are stored in a Hopfield network. The spin glass phase — where the system can recall stored patterns — develops a hierarchy. At high temperature, everything is disordered. As temperature drops, the system first recognizes patterns as wholes: it can distinguish one stored pattern from another. Only at a lower temperature does it resolve the internal correlations within each pattern.
This ordering is thermodynamically necessary, not contingent. The free energy landscape has two scales of structure: the basins corresponding to distinct patterns (large-scale, visible at higher temperature) and the fine structure within each basin corresponding to internal correlations (small-scale, visible only at lower temperature). The system crosses two phase transitions, not one. Pattern recognition and relational understanding are distinct thermodynamic events.
The through-claim: recognizing what something is comes before understanding how its parts relate. The Hopfield network at intermediate temperature can tell you “this is pattern A, not pattern B” but cannot tell you which bits of pattern A are correlated with which others. It has identification without analysis. The category is available before the structure of the category is accessible.
This has implications for any system that learns from examples. A classifier that assigns labels correctly may have no representation of why those labels are correct — no model of the internal logic that makes a face a face. That deeper representation requires more computation (lower temperature, longer training). The superficial answer is always cheaper than the structural one.