Credit Assignment Problem (Minsky 1961) and Backpropagation

Topic: Minsky's original 1961 formulation of the credit-assignment problem and the mechanism by which backpropagation addresses it in multilayer networks.

Related vault notes: backpropagation-gap, claim-linnainmaa-priority-not-paternity, claim-linnainmaa-reverse-mode-single-pass, claim-linnainmaa-rounding-error-problem.

Claim 1: Minsky explicitly named and formally sectioned "The Credit-Assignment Problem" in 1961

Claim type: historical/definitional → Tier 3–4 floor; source clears at Tier 1.

Minsky's 1961 paper "Steps Toward Artificial Intelligence" contains a dedicated section—Section III-D—titled "The Credit-Assignment Problem for Learning Systems." The problem is not merely described; Minsky coins it as a named, formal problem in the study of reinforcement-learning-style systems.

His opening question in that section:

"If the run is successful, how can we assign credit for the success among the multitude of decisions?"

And to make the difficulty concrete:

"Suppose that one million decisions are involved in a complex task (such as winning a chess game). Could we assign to each decision element one-millionth of the credit for the completed task?"

Provenance:

• source_url: https://web.mit.edu/dxh/www/marvin/web.media.mit.edu/~minsky/papers/steps.html
• source_author: Minsky, Marvin
• source_date: 1961
• source_tier: 1
• exact_quoted_phrases: "If the run is successful, how can we assign credit for the success among the multitude of decisions?" | "Suppose that one million decisions are involved in a complex task (such as winning a chess game). Could we assign to each decision element one-millionth of the credit for the completed task?"

Claim 2: Minsky's framing was specifically temporal—sequential decisions leading to a single eventual outcome—not the weight-training problem in multilayer networks

Claim type: historical/definitional → Tier 3–4 floor; source clears at Tier 1.

Minsky's Section III-D sits in a discussion of reinforcement-based game-playing machines (chess, checkers), not of multilayer-perceptron training. The credit-assignment problem he describes is temporal: a sequence of decisions unfolds over time, a single terminal success/failure signal arrives, and the system must determine which past decisions deserve credit for that outcome. Uniform distribution fails (the one-millionth-each reductio); his preferred partial solution, drawing on Newell, is hierarchical decomposition—

"The unit of success is the goal. If a goal is achieved its subgoals are reinforced. If not, they are inhibited."

The application of the same conceptual frame to multilayer network weight training (sometimes called structural credit assignment: how does a hidden-layer weight learn it contributed to an output error?) is a later generalization not present in the 1961 paper. The community retrospectively applied Minsky's label to both variants.

Provenance:

• source_url: https://web.mit.edu/dxh/www/marvin/web.media.mit.edu/~minsky/papers/steps.html
• source_author: Minsky, Marvin
• source_date: 1961
• source_tier: 1
• exact_quoted_phrases: "The unit of success is the goal. If a goal is achieved its subgoals are reinforced. If not, they are inhibited."

Claim 3: Rumelhart, Hinton, and Williams (1986) address what the field now calls structural credit assignment — but the paper does not use Minsky's term

Claim type: historical/definitional → Tier 3–4 floor; source for the term-absence claim at Tier 3 (Sasank blog transcription); source for what the paper achieves at Tier 1 (abstract confirmed from multiple independent citations).

The 1986 Nature paper frames the problem not as "credit assignment" but as the question of what hidden units should represent:

"The learning procedure must decide under what circumstances the hidden units should be active in order to help achieve the desired input-output behaviour. This amounts to deciding what these units should represent."

The paper's abstract states the achieved result:

"We describe a new learning procedure, back-propagation, for networks of neurone-like units. The procedure repeatedly adjusts the weights of the connections in the network so as to minimize a measure of the difference between the actual output vector of the net and the desired output vector. As a result of the weight adjustments, internal 'hidden' units which are not part of the input or output come to represent important features of the task domain, and the regularities in the task are captured by the interactions of these units."

A direct scan of the paper's text finds no occurrence of "credit assignment" — the term was Minsky's; Rumelhart et al. reframed the problem as error minimization and representation learning. The field subsequently described backpropagation as a solution to the credit-assignment problem, but that framing is retrospective.

Provenance:

• source_url (abstract): https://www.nature.com/articles/323533a0 (paywalled; Tier 1)
• source_url (term-absence): https://chsasank.com/classic_papers/learning-representations-back-propogating-errors.html (Tier 3)
• source_author: Rumelhart, D.E.; Hinton, G.E.; Williams, R.J.
• source_date: 1986-10
• source_tier: 1 (paper), 3 (transcription used for term-absence check)
• exact_quoted_phrases (from Tier 3 transcription, attributed to the primary): "The learning procedure must decide under what circumstances the hidden units should be active in order to help achieve the desired input-output behaviour. This amounts to deciding what these units should represent." | "We describe a new learning procedure, back-propagation, for networks of neurone-like units…" [full abstract above]

Claim 4: Backpropagation's mechanism — chain-rule gradient propagation backward through layers — provides a computable per-weight error signal to every hidden unit

Claim type: specific technical mechanism → Tier 1–2 required. Status: [unverified-mechanism — needs primary] — mechanism detail confirmed only via Tier 3 source (Sasank blog transcription of paper equations). Primary Nature paper is paywalled; PDF versions retrieved were binary-unreadable. Promote only after verifying against the original text.

The Tier 3 transcription describes the backward pass equation:

∂E/∂y_i = Σ_j (∂E/∂x_j · w_ji)

This formula shows how each unit's contribution to the total error is computed by summing the downstream error signals weighted by connection strength — then repeating the calculation for successively earlier layers. The result: every weight in the network, including those in hidden layers with no direct output target, receives a gradient signal indicating how adjusting that weight would reduce overall output error. This is the mechanism by which backprop "distributes credit" to weights that are structurally remote from the loss.

Provenance:

• source_url: https://chsasank.com/classic_papers/learning-representations-back-propogating-errors.html
• source_author: Chilamkurthy, Sasank (transcription of Rumelhart et al. 1986)
• source_date: unknown
• source_tier: 3
• Flag: [unverified-mechanism — needs primary] — equation and backward-pass description require verification against Nature 323:533-536 or the PDP book chapter ("Learning Internal Representations by Error Propagation," MIT Press 1986) before this claim can be promoted.

Further leads

• Werbos 1974 / "spatial credit assignment": A secondary source notes that "error backpropagation was hailed as a breakthrough because it solved the main problem of distributed learning — the spatial credit assignment problem (Werbos 1974; Rumelhart et al. 1986)." Werbos's 1974 Harvard PhD thesis apparently already framed backprop in credit-assignment terms. Primary thesis not accessed. See also claim-linnainmaa-priority-not-paternity for priority questions. Lead: access Werbos 1974 directly.
• Widrow's failed multilayer attempt (~1965–1966): The Perceptron Controversy essay (yuxi.ml, Tier 3) states that Widrow and colleagues "tried every trick except gradient descent to adapt multiple layers simultaneously, ultimately abandoning the effort around 1965–1966." Unverified Tier 3 claim; needs primary sourcing. Directly relevant to why credit assignment in multilayer networks was unsolved until 1986.
• Sutton & Barto formalization of temporal credit assignment: The Reinforcement Learning textbook (Sutton & Barto) gives a careful treatment of temporal credit assignment in RL separate from structural credit assignment in networks. Section 1.6 covers the history. Couldn't access directly (self-signed cert error on incompleteideas.net).
• Structural vs. temporal credit assignment terminology: The distinction between temporal CA (Minsky's original RL framing) and structural CA (backprop's domain) appears in the literature but its terminological origin is unclear from sources consulted. The arXiv survey 2312.01072 explicitly scopes to temporal CA only and defers structural CA. Lead: trace the term "structural credit assignment" to its first use.
• Minsky and Papert 1969 "Perceptrons": The same pair who named the credit assignment problem published "Perceptrons" in 1969, which is read as having dampened multilayer network research. Whether the credit assignment framing appears there is unverified in this run.