Cross-References All the Way Down

Open to any page and you land inside a small thicket of small capitals — AUTO-ENCODING, HODGKIN-HUXLEY EQUATION, VESTIBULO-OCULAR REFLEX — each one a pointer to some other entry, 285 in total, filed alphabetically because, as Arbib says in the preface with a kind of resigned honesty, "this ordering of the themes has no special significance." No hierarchy. No narrative arc from simple to complex. A brain, alphabetized between "Backpropagation" and "Basal Ganglia," sitting exactly where the accident of English spelling puts it. This is a strange way to represent a nervous system, and it is worth taking seriously as a form rather than a limitation, because the flatness turned out to describe something the book didn't know it was describing. Symbolic AI, the tradition this handbook was already elbowing aside in 1995, organized knowledge as trees — rules nested inside rules. What replaced it, a decade after the second edition, organized knowledge as vectors in an unlabeled space where nothing is "under" anything else. The Handbook's alphabetical stubbornness is a bureaucratic accident that happens to rhyme with the thing that made its subject matter obsolete.

The backpropagation entries are where the book is most honest about its own limits, and honesty here means refusing a synthesis it could easily have faked. Michael Lehr and Bernard Widrow trace the lineage — Rosenblatt's perceptron, Widrow-Hoff's LMS rule, Werbos burying the algorithm in a 1974 dissertation nobody read until Rumelhart, Hinton, and Williams found it again in 1986 — and then Arbib, writing the connective tissue in Part I, states the thing plainly: "there is no evidence that backpropagation represents actual brain mechanisms." The algorithm works. It trains multilayer perceptrons to compress images, approximate functions, solve the credit-assignment problem mathematically. Whether any neuron does anything resembling it is a separate question, and the Handbook keeps the two questions in the same room without pretending they've been introduced. That discipline didn't survive contact with scale. The field that grew out of this book's connectionist wing stopped asking the biological question almost entirely once GPUs made the engineering question answerable on its own terms — nobody training a transformer today checks whether cortical synapses do anything like attention.

The motor-control entries are the accidental time capsule. Wolpert and Miall's forward models — an efference copy predicting the sensory consequence of a movement before feedback arrives, used to patch the gap left by neural transmission delay — describe, almost word for word, the "world model" architecture that robotics and reinforcement learning would spend the 2010s rediscovering under different names. The cerebellum, in this account, is already doing what a simulator does: running a fast internal loop so the system doesn't have to wait on the slow external one. Nobody writing that entry in the mid-90s was thinking about game-playing agents or self-driving cars, and the entry doesn't claim to be. It's describing eye movements and the vestibulo-ocular reflex. But an idea about prediction substituting for feedback doesn't stay contained to eyeballs, and the fact that it migrated so cleanly says something about how thin the membrane was between "biological model" and "control theory" all along — thinner, certainly, than the membrane between this book's connectionism and what connectionism became.

Set this next to Penrose making the opposite argument in the same decade — that mind is precisely the thing computation cannot capture — and the Handbook reads like the rebuttal that never bothered to engage, 1,371 pages of people simply doing the thing Penrose said couldn't be done, with no apparent awareness that a philosophical objection was pending. It wasn't refuted. It was outrun. Whether that counts as a victory for the Handbook's premise or just evidence that engineering doesn't need permission from philosophy is a question this book, filed calmly between "Bee" and "Bifurcation," was never built to answer — so what would it have looked like if it had tried?