No Single “True” Decomposition Exists

The aspiration to “carve neural networks at their joints” imports a problematic assumption: that complex systems have a unique and natural decomposition, independent of explanatory context.

Philosophers of science have long emphasized that mechanisms span multiple levels of organization. In biology, scientists study whole ecosystems, individual organisms, organ systems, cellular mechanisms, molecular interactions — none of these levels has a unique claim to being “the real” level. Mechanisms are nested within mechanisms.

The practical implication: Which level of mechanism is most important depends on the pragmatic goals of researchers. A model inference results from a complex web of small routines that have proven adaptive across tasks. Different decompositions offer partial but contextually salient insights.

This supports explanatory pluralism — scientific understanding often requires integrating multiple, non-reducible models, tailored to different explanatory aims. Different decompositions may be more or less useful depending on whether we’re trying to control outputs, understand generalization, or detect deception.

Decompositions should be evaluated not by how well they mirror the “real” structure, but by how effectively they support causal understanding, prediction, and intervention across different research contexts.

Related: [None yet]