The Systematicity Argument

Definition

Fodor and Pylyshyn observe that human cognitive abilities are never "punctate" — nobody can think one thought without being able to think structurally related ones: anyone who can infer John and Mary went to the store from John and Mary and Susan went to the store can also infer John went to the store from John and Mary went to the store, and so on across countless parallel cases. They argue only an architecture with combinatorial syntax and semantics — a classical symbol system — can explain why this pattern is a near-truism rather than an accident, because a connectionist network can be built to exhibit it but has no internal mechanism that forces it to.

In the Book

The chapter's core move is to show that "connectionist architecture tolerates gaps in cognitive capacities" — since connectionist representations are, structurally, just a list of causally-connected nodes, and "lists, qua lists, have no structure; any collection of items is a possible list." A network could in principle have a node for aRb without one for bRa, producing a mind that could think one relational thought but not its systematic partner — yet no such minds are found. Fodor and Pylyshyn anticipate the reply that a connectionist could simply build networks that happen to respect systematicity, and reject it: stipulating the pattern isn't explaining it; you need a mechanism that enforces it, and "the only mechanism that is known to be able to produce pervasive systematicity is classical architecture," which requires exactly the internally-structured, syntactically-combinable representations that connectionism, by its own principles, forgoes. The chapter also surveys and rebuts the standard pro-connectionist arguments (the "hundred-step constraint" on neural speed, graceful degradation under damage, and difficulty explaining tacit/nonverbal skill) as motivating parallelism at the implementation level without undermining classical symbolic structure at the level of cognitive architecture.

Why It Matters

The systematicity argument gives a general test for any proposed cognitive or computational architecture: does it merely permit an observed regularity, or does its structure actually force that regularity to hold? A model that can be tuned to fit a pattern after the fact explains less than one whose architecture makes the pattern unavoidable — a distinction useful whenever you're evaluating whether a proposed mechanism is doing real explanatory work or just being retrofitted to match the data.