Negative Transfer

Negative transfer occurs when knowledge or skills acquired in one context impair performance in a different context. Unlike positive transfer — the typical goal of transfer learning — negative transfer arises when the structural assumptions of the source domain actively mislead the learner in the target domain. A model trained to recognize animals in photographs may fail catastrophically when transferred to medical imaging if the texture statistics it learned correlate with disease in the training set but not in the hospital. Negative transfer is the empirical signature of an unexamined assumption: that similarity between domains is sufficient for knowledge reuse. It is not. Similarity must be structural, not merely superficial.

The phenomenon appears in human cognition as well. A pianist learning the accordion may struggle because the similar keyboard layout encodes incompatible motor mappings. In institutional contexts, a management practice that succeeded in one company may destroy another because the causal structures of the two organizations differ. Negative transfer reveals that the universe's grammar is not uniformly readable — some chapters share syntax but contradict in semantics.

Negative transfer is closely related to the broader problem of distribution shift — the mismatch between training and deployment distributions that underlies most failures of machine learning in the wild.