Jürgen Schmidhuber

• 1. Introduction

  ›

  Jürgen Schmidhuber was born on the 17th of January 1963, and for decades the New York Times ran a profile of him under the headline "When A.I. Matures, It May Call Jürgen Schmidhuber 'Dad'." That headline captures something both admiring and unresolved about his story. Here is a man whose research touches virtually every corner of modern artificial intelligence, who leads institutions in Switzerland and Saudi Arabia simultaneously, and who is equally known for his landmark technical breakthroughs and for the fierce public battles he wages over who deserves credit for them. How did a German computer scientist working in the relative quiet of Lugano end up at the center of one of the most consequential intellectual disputes in the history of technology? That question runs through everything that follows.

• 2. The Making of an AI Pioneer

  ›

  Schmidhuber completed his undergraduate degree in 1987 and his doctorate in 1991, both at the Technical University of Munich. His PhD advisors were Wilfried Brauer and Klaus Schulten. By 1995 he had taken the helm of IDSIA, the Dalle Molle Institute for Artificial Intelligence Research in Switzerland, a position he still holds. He taught at Munich from 2004 until 2009, then moved to the Università della Svizzera Italiana in Lugano, where he was a professor of artificial intelligence until 2021. Since 2021 he has also directed the AI Initiative at the King Abdullah University of Science and Technology in Saudi Arabia, known as KAUST. The thread running through these appointments is a conviction he has carried since the 1970s: that intelligent machines could learn, improve on their own, and become smarter than their creator within a single lifetime.

• 3. The LSTM Breakthrough

  ›

  Sepp Hochreiter was one of Schmidhuber's students, and in 1991 Schmidhuber supervised his diploma thesis, which Schmidhuber considered "one of the most important documents in the history of machine learning." That thesis tackled the vanishing gradient problem, a flaw that caused neural networks to forget relevant information when sequences grew long. The work seeded the creation of long short-term memory, or LSTM, a type of recurrent neural network. The name LSTM itself did not appear until a 1995 technical report, and the architecture's defining publication came in 1997, co-authored by Hochreiter and Schmidhuber. The standard LSTM architecture followed in 2000, credited to Felix Gers, Schmidhuber, and Fred Cummins. A further refinement, the so-called "vanilla LSTM" using backpropagation through time, was published in 2005 alongside his student Alex Graves, with the connectionist temporal classification training algorithm arriving in 2006. CTC was then applied to end-to-end speech recognition. Throughout the 2010s, LSTM was the dominant technique for natural language processing tasks across both research and commercial applications.

• 4. Before GANs and Transformers

  ›

  In 1991, Schmidhuber published a framework describing two adversarial neural networks contesting each other in a zero-sum game. One network generated patterns; the other learned to predict the environment's reactions to those patterns. He called this "artificial curiosity." When the generative adversarial network, or GAN, was introduced in 2014, Schmidhuber described it as a special case of artificial curiosity in which the environmental reaction is a simple binary signal. That same year, highway networks arrived, developed by Rupesh Kumar Srivastava, Klaus Greff, and Schmidhuber using LSTM principles, enabling feedforward networks with hundreds of layers at a time when state-of-the-art models topped out at 20 to 30. The residual neural network, or ResNet, published in December 2015, is itself a variant of the highway network. In 1992, Schmidhuber also published the fast weights programmer, which was later shown to be equivalent to the unnormalized linear transformer, a precursor of the architecture now synonymous with large language models.

• 5. Vision at Scale

  ›

  Dan Ciresan joined IDSIA as a postdoc, and in 2011 his team with Schmidhuber achieved CNNs running on graphics processing units that were 60 times faster than an equivalent CPU implementation. For context, an earlier CNN-on-GPU by Chellapilla and colleagues from 2006 had managed only a fourfold speedup. The IDSIA team's deep CNN achieved the first superhuman performance in a computer vision contest in August 2011. Between the 15th of May 2011 and the 10th of September 2012, those CNNs won four more image competitions and improved the state of the art across multiple benchmarks. The approach became central to the entire field of computer vision. The underlying CNN designs drew on architectures introduced much earlier by Kunihiko Fukushima, a lineage that Schmidhuber has consistently traced and cited.

• 6. Credit Disputes and the Turing Award

  ›

  Geoffrey Hinton, Yoshua Bengio, and Yann LeCun shared the 2018 Turing Award for their work in deep learning. Schmidhuber wrote a widely read 2015 article, described as "scathing," arguing that the three "heavily cite each other" but "fail to credit the pioneers of the field." In a statement to the New York Times, LeCun responded that Schmidhuber was "manically obsessed with recognition" and claimed credit he did not deserve, accusing him of challenging the originality of presented work at the end of talks. Schmidhuber replied that LeCun made those accusations without justification and without a single supporting example, and he published detailed accounts of specific priority disputes with all three researchers. The AI community coined the verb "schmidhubered" to describe his habit of publicly contesting originality, a practice that some younger researchers view as a rite of passage and that others see as evidence that his genuine accomplishments have been underappreciated because of his confrontational style.

• 7. NNAISENSE and Future Goals

  ›

  In 2014, Schmidhuber formed a company called NNAISENSE to pursue commercial applications of AI in finance, heavy industry, and self-driving cars. Sepp Hochreiter, Jaan Tallinn, and Marcus Hutter serve as advisers. Sales were under eleven million US dollars in 2016, though Schmidhuber stated at the time that the emphasis was on research rather than revenue. NNAISENSE raised its first round of capital funding in January 2017. Schmidhuber's original goal for the company was to build a general-purpose AI by training a single system sequentially on a variety of narrow tasks, but as of 2026 he has said that NNAISENSE's focus has shifted from artificial general intelligence to asset management. His personal ambitions, meanwhile, remain very large: he differentiates between tool AI, designed to improve specific human outcomes like healthcare, and autonomous AI that sets its own goals, conducts its own research, and eventually explores the universe. He has received the Helmholtz Award of the International Neural Network Society, in 2013, and the Neural Networks Pioneer Award of the IEEE Computational Intelligence Society in 2016, cited for "pioneering contributions to deep learning and neural networks," and he is a member of the European Academy of Sciences and Arts.