Paul Werbos

Life and career

The algorithm that trains every deep network on earth was written down in 1974, in a Harvard doctoral thesis about forecasting social and political behavior, and almost nobody read it.

Paul Werbos, born in 1947, was not a computer scientist by training or temperament. He studied at Harvard and at the London School of Economics, and his doctoral work sat inside an interdisciplinary problem: how do you build quantitative models of human and social systems that actually predict? His advisor's world was political science and social forecasting, and the mathematical question that came out of it was general — given a layered, nonlinear model with many parameters and a target you want to match, how do you compute all the sensitivities efficiently enough to fit it?

His 1974 dissertation, *Beyond Regression: New Tools for Prediction and Analysis in the Behavioral Sciences*, answered that question with what he called the method of ordered derivatives: propagate the chain rule backward through the computation in reverse topological order, so that the cost of computing the gradient with respect to *all* parameters is a small constant multiple of the cost of one forward pass. Applied to a multilayer network of nonlinear units, that is backpropagation, complete. It is worth saying plainly that Werbos was not the only person in the vicinity — Seppo Linnainmaa had published the reverse-mode accumulation of derivatives in 1970 in the numerical-analysis literature, and the general idea of reverse-mode differentiation has several roots — but Werbos's contribution was to derive it *for* the training of layered nonlinear models and to argue explicitly that this was the way to fit them.

Werbos has described the inspiration in unusually personal terms, tracing his intuition that credit and blame should flow backward through a system to ideas he had absorbed about how the mind assigns significance. The route into the algorithm mattered less than the timing, which could not have been worse. He arrived at the answer during the coldest stretch of the first AI winter, in a field that had been told by an influential 1969 result that multilayer networks were the interesting case and had no way to train them. There was no community waiting for the fix. His thesis sat unread for a decade, until Rumelhart, Hinton and Williams made the same idea an event in 1986 and the field discovered, somewhat awkwardly, that it had been in the library the whole time.

The rest of his career took a shape that people often miss. He worked in energy forecasting for the U.S. government, then joined the National Science Foundation in 1989 as a program director, where for roughly a quarter century he directed federal funding into neural networks, adaptive control, and what he called intelligent systems. During the long stretch when neural network research was unfashionable, Werbos was one of the people quietly keeping money flowing to it. He republished the thesis in 1994 as *The Roots of Backpropagation*, partly to make the record accessible, and he received the neural network community's pioneer recognition. He remains an unusual figure: the man with the strongest priority claim in modern machine learning, who spent his working life as a program officer rather than a professor.

Key contributions

**Ordered derivatives and reverse-mode credit assignment.** The formal object Werbos introduced is the ordered derivative — the total sensitivity of an output to an intermediate quantity, accounting for every downstream path through which it acts, computed by traversing the computation graph in reverse. For a network with *L* layers, the forward pass computes activations; the backward pass computes ∂E/∂net_j at each unit by summing the contributions from the units it feeds, multiplied by the local derivative. The cost is O(number of edges), the same as the forward pass — this is the entire reason deep learning is computationally possible. A naive forward-mode computation of the same gradients costs a factor of *P* more, where *P* is the parameter count, which at modern scale is a factor of billions. Werbos understood and stated this efficiency argument; it is not a detail of the implementation, it is the theorem.

**Backpropagation through time.** Werbos generalized the method to recurrent systems by unrolling the recurrence across time steps into a deep feedforward graph with shared weights, then applying reverse-mode accumulation to the unrolled graph. His 1990 *Proceedings of the IEEE* paper on BPTT is the standard reference, and BPTT remains the training algorithm for every recurrent architecture in use. It is also where the pathology lives: unrolling a recurrence multiplies the same Jacobian repeatedly, so gradients decay or explode geometrically in the horizon length — the vanishing gradient problem, whose mechanism is a direct consequence of the algorithm Werbos wrote down.

**Adaptive dynamic programming and adaptive critics.** The contribution he himself considers most important is the least famous. Exact dynamic programming is intractable in continuous, high-dimensional state spaces. Werbos proposed a family of designs in which a differentiable function approximator learns the value function (or its gradient) and a second approximator learns the policy, with the two trained against each other — Heuristic Dynamic Programming, Dual Heuristic Programming, and their globalized variants. In DHP the critic learns ∂J/∂x, the *derivative* of the value function, rather than J itself, which is precisely the quantity a controller needs and which avoids differentiating a noisy scalar estimate. This is the actor-critic architecture in its general form, developed for control rather than for games, and it is the ancestor of a great deal of modern deep reinforcement learning.

In battle

Werbos plays like a man holding one very important key and few other tools: mean 20.5, median 15, sixty-nine problems at or below 20, and only two above 80. He is a low-baseline, high-spike character, and both spikes are the same idea.

He peaks at 88 on "XOR," and the game's reasoning is worth quoting in spirit: the 1969 result said single layers cannot represent parity, the open question was how to train multilayer networks, and Werbos answered it in writing five years into the winter and twelve years before the field noticed. The game docks him for two honest reasons — he did not publicly demonstrate the fix on XOR at the time, and the representational insight was someone else's — which is why this is 88 rather than 99. His second dominant showing, "The gradient that vanishes" at 85, is the same claim from the other direction: he wrote the algorithm whose repeated Jacobian products *create* the pathology, and his BPTT paper is the standard analysis of what unrolling does. "The million-parameter valley" (68) and "Sixty thousand digits" (58) follow directly: he supplies the gradient machinery that makes both tractable, without having built either system.

His most distinctive comparative strength is reinforcement learning, where he averages 47.0 — the highest of any category on his sheet, and higher than most of the deep-learning founders around him. "Climb the yield surface" (50), "The two slot machines" (48), and "Learn the optimum while acting otherwise" (46) are all adaptive-critic territory: sequential decision-making with function approximation, which is exactly what he spent thirty years building. A student who assumes Werbos is a pure backprop specialist will be surprised by these.

The losses expose what the ordered-derivative hammer cannot hit. "Just look at the neighbors" at 3 is the floor, and the explanation is exact: nearest-neighbor asymptotics and the curse of dimensionality are a nonparametric theory that engages nothing in his record. Classical regression diagnostics defeat him — "The coefficient that flips sign" (4) is confounding and multicollinearity, "Why tall fathers have shorter sons" (5) is regression to the mean — because a general-purpose function fitter has no theory of *why* a coefficient means anything. "The spam in the inbox" (5) rewards a generative probabilistic model, not a discriminatively fitted one. "What is a bit, exactly?" (5) is information theory, a foundational framework he uses but did not build. And "Cut through the interior" (6) is instructive: interior-point methods are convex optimization with polynomial-time guarantees, whereas Werbos's whole contribution is a way of doing gradient descent on non-convex problems while making no guarantees at all.

Play Werbos when the problem is *credit assignment through a deep or temporal structure*, or sequential control with a learned value function. Bench him whenever the question is what the fitted numbers mean.