'Credit assignment' names both the sociology of scientific credit and backpropagation's technical problem — and Schmidhuber sits on the hinge

The phrase "credit assignment" carries two nearly unrelated meanings that converge on the history of AI. In the sociology of science, credit assignment is the (mis)allocation of recognition for a discovery — the subject of the Merton–Stigler tradition. In machine learning, "the credit-assignment problem" is the technical question of how to apportion responsibility for an outcome across many decisions or weights, named by Minsky in 1961 (claim-minsky-1961-named-credit-assignment); backpropagation is the method most associated with solving its structural form (myth-structural-credit-assignment-from-minsky-1961).

Jürgen Schmidhuber occupies the hinge between the two senses. He opens his Annotated History of Modern AI and Deep Learning (arXiv:2212.11279) with the technical definition — "Machine learning (ML) is the science of credit assignment" — while the same document runs the field's most sustained campaign in the sociological sense: reassigning priority for deep-learning ideas from famous names to earlier, more obscure ones. The genealogy he builds (Linnainmaa → Dreyfus → Werbos → …) is itself an exercise in reversing misassigned credit — the move examined at claim-dreyfus-1973-lineage-link-uncorroborated and claim-werbos-1974-no-credit-assignment-language.

That genealogical move — reducing an apparent breakthrough to descent from prior work — is the same one Tertius Chandler states as a general law (claim-chandler-simultaneous-discoveries-are-incremental-repackagings), applied here to backpropagation. It stands against the Ogburn–Thomas/Merton reading of multiple discovery, in which convergent inventions are genuinely independent and "in the air." So the two senses of credit assignment are not merely a pun: Schmidhuber's technical subject (apportioning credit inside a network) and his historiographic method (apportioning credit across a field) are the same operation at two scales. See moc-backpropagation-origins.