Simon argued twice (1993, 1995) that connectionist neural networks qualify as physical symbol systems under his own definition, rather than counterexamples refuting the hypothesis

Simon engaged connectionism directly, and repeatedly, late in his career — not by refuting it, but by arguing it belongs inside his own framework rather than outside it. In "Situated Action: Reply to Reviewers" (Vera & Simon, Cognitive Science, 1993), responding to critics of physical-symbol-system cognitive theory, the authors discuss the "Navlab" road-steering connectionist network at length and conclude that its trained weight-pattern — which "represent[s] (denote[s]) different curvatures of the road" and is then read out to steer the wheel — "satisfies our definition of symbol." Addressing the dispute directly: "This does not, however, make it a nonsymbolic system."

Two years later, in "Artificial Intelligence: An Empirical Science" (Artificial Intelligence, 1995, pp. 104–105; source_url: https://ic.unicamp.br/~wainer/cursos/2s2006/epistemico/simon-ia.pdf), Simon named the dispute explicitly and staked the same ground alone:

"If we define symbols (as I have, above) as patterns that denote, then connectionist systems and Brooks' robots qualify as physical symbol systems."

Read together, these are not two separate claims but one sustained theoretical move, evidenced twice: rather than concede that connectionist networks are a rival paradigm refuting the physical symbol system hypothesis, Simon redefined "symbol" broadly enough to absorb them as a species of symbol system. See moc-backpropagation-origins.