Jurgen Schmidhuber

Life and career

There is a version of the last forty years of machine learning in which the central lab is not in Toronto, New York or Montreal but in a small institute above Lake Lugano, and its director has spent three decades trying to get the rest of the field to read it.

Jürgen Schmidhuber was born in Munich in 1963 and did both his diploma and his doctorate at the Technical University of Munich, finishing the PhD in 1991. That single year is remarkable in retrospect: working in Munich, he and the students around him produced the vanishing-gradient diagnosis, the first hierarchical recurrent architecture trained by self-supervised prediction, an adversarial two-network scheme, and the formal proposal that an agent should generate its own reward from the improvement of its world model. Any one of these would be a strong career. Almost none of them were noticed at the time.

In 1995 he became a director of IDSIA, the Dalle Molle Institute for Artificial Intelligence in Lugano, Switzerland, and stayed for a quarter century, building a group with a distinctive profile: small, theoretically ambitious, and consistently early. LSTM came out of it in 1997, with Sepp Hochreiter. Connectionist temporal classification, the loss function that made end-to-end speech and handwriting recognition possible without frame-level alignment, came out of it in 2006 with Alex Graves. In 2011 and 2012 his group's GPU-trained convolutional networks won a series of international competitions — traffic sign recognition at superhuman accuracy, Chinese handwriting, medical image analysis — in the same window that AlexNet was being built in Toronto. In 2015 his students introduced highway networks, which use learned gates to let signal pass through very deep stacks unimpeded, immediately preceding and closely related to the residual connections that became universal. He co-founded the company NNAISENSE and later took a leadership role in the AI initiative at KAUST in Saudi Arabia. Marcus Hutter, who developed the AIXI theory of universal intelligence, was his doctoral student.

He is also, unavoidably, the field's most persistent archivist of its own credit assignment. Schmidhuber has publicly and repeatedly contested the attribution of major deep learning results — including in a widely circulated critique of a 2015 survey by Hinton, LeCun and Bengio, and in an exchange with Ian Goodfellow at a NIPS tutorial over the relationship between generative adversarial networks and his 1990s predictability-minimization work. He maintains detailed annotated histories arguing for earlier priority on a long list of ideas. The disputes are real, documented, and unresolved; different senior researchers assess them differently, and the point here is not to adjudicate but to note that this campaign is as much a part of his public identity as the LSTM is. It has also, arguably, cost him: the sociology of a field determines whose name attaches to an idea, and he has been better at inventing than at that.

Key contributions

**Long short-term memory (1997, with Hochreiter).** The constant error carousel is the core idea: a memory cell with a self-recurrent connection of weight exactly one, so that its internal state carries forward unchanged and gradients flow backward through it undiminished across arbitrarily many steps. Around this sits a set of learned multiplicative gates — input, output, and later the forget gate added by Gers, Schmidhuber and Cummins in 2000 — that decide what enters the cell, what leaves it, and when it should be cleared. The architecture converts the vanishing-gradient problem from a property of the dynamics into something the network can *control*, and it carried sequence modeling in speech recognition, machine translation, and handwriting for nearly two decades.

**Artificial curiosity and intrinsic motivation (1990–91 onward).** Schmidhuber proposed that an agent be rewarded not by the environment but by *the improvement in its own predictive model* — the reduction in the error, or equivalently the gain in compression, that results from an observation. This is a formal answer to exploration in sparse-reward environments, and his later "formal theory of fun and creativity" made the argument explicit: reward learning progress, not surprise. The distinction matters enormously in practice, because rewarding raw prediction error produces an agent that will stare forever at an unpredictable but uninformative source of noise, while rewarding *progress* will not.

**Predictability minimization and adversarial training (1992).** Two networks with opposed objectives — one trying to make a representation's components mutually unpredictable, the other trying to predict them — is the structure Schmidhuber cites as an early adversarial learning scheme. The relationship to modern GANs is precisely what is disputed; the mechanism is real.

**Fast weight programmers and history compression (1991–92).** One network generates, on the fly, the weights of another via outer-product updates to a rapidly changing weight matrix; a separate strand compressed a sequence's history by having a higher-level recurrent network model only the events its lower level failed to predict. The first has been shown to be closely related to linearized attention; the second is an early instance of self-supervised hierarchical sequence pretraining.

**Connectionist temporal classification and deep GPU vision.** CTC marginalizes over all alignments between an input sequence and a shorter label sequence via dynamic programming, eliminating the need for frame-level labels — the reason end-to-end speech recognition works. The 2011–12 competition wins demonstrated that deep convolutional networks on GPUs were already superhuman on constrained recognition tasks.

**Theoretical universal AI.** The Gödel machine, a self-referential agent that rewrites its own code only upon proving the rewrite beneficial; the speed prior; and the optimal ordered problem solver. These are formal rather than practical, and they mark the more speculative wing of his program.

In battle

Schmidhuber's numbers — mean 25.2, median 20, five problems above 80, fifty-three at or below 20 — describe a specialist with a narrow but very deep band, and the band is *sequences and self-directed agents*.

His two peaks sit at 96. "The room with no reward" is the sparse-reward exploration problem, and the game's explanation credits him with the specific mechanism it needs: self-generated intrinsic reward from world-model improvement, proposed in 1990–91, together with the compression-progress framing that anticipates and avoids the noisy-TV failure mode. His reinforcement learning category average of 73.0 is the highest single category on his sheet by a wide margin, which will surprise students who file him as an LSTM person. "The gradient that vanishes" is the other 96 — his lab's diagnosis and his lab's cure, shared with Hochreiter, who edges him on it.

The sequence cluster follows: "The sentence in a single vector" (85), because LSTM is the encoder whose bottleneck the problem describes; "Attention replaces recurrence" (80), an unusually generous and defensible score given his fast-weight work and the fact that his architecture is the incumbent being replaced; "Sixty thousand digits" (85) and "The tilted scan" (64), which rest on his group's competition-winning GPU convolutional networks rather than on any recurrent claim. His computability average of 31.5 — high for a deep learning figure — comes from the Gödel machine and universal-learner side of his work, which almost nobody else on the roster can touch.

The losses are unusually clean. Optimization crushes him: "The variables that must be whole" (5), "Relax, then round" (5), "The corner that sparsity loves" (4), "Optimize across the datacenter" (4) — integer programming, convex relaxation, ℓ1 geometry and distributed convex optimization, giving him an optimization category average of 14.5. His entire methodology is architectural: build a structure that makes the learning problem tractable and run SGD on it. That produces no theory of relaxations, certificates, or sparsity geometry.

Classical statistics and causality are worse. His causality average of 5.0 is his floor, and "The paradox in the admissions data" at 4 shows why — Simpson's paradox is a question about what a conditional distribution *means*, and no amount of universal-learner ambition addresses it. "Three species of iris" at 3 draws the game's driest line: universal learners are a strange tool to bring to four flower measurements.

Play Schmidhuber when the problem involves memory across time, an agent that must decide what to explore, or a deep stack that needs a path for signal to survive. Bench him when the problem is small, classical, convex, or about interpretation rather than construction.