IDSIA: artificial intelligence in Lugano since 1988

Angelo Dalle Molle and his institutes

The oldest of the research institutions described in this chapter was founded by a private benefactor eight years before the canton had a university. Angelo Dalle Molle (1908– 2002) was an Italian entrepreneur from the Veneto who had made his fortune with the artichoke liqueur Cynar and who devoted a large part of it to the promotion of research on the quality of life and on the relationship between people and machines.¹ He established a trust in Switzerland to fund research outside traditional university structures and created three institutes: IDSIA in Lugano, IDIAP in Martigny and ISSCo in Geneva. All three are today integrated into their local academic institutions.² The Istituto Dalle Molle di studi sull'intelligenza artificiale (IDSIA, Dalle Molle Institute for Artificial Intelligence) was founded in Lugano in 1988. Dalle Molle, in the words of the institute, "envisioned a world in which technological progress and human development would jointly enhance quality of life".³ At the celebration of the institute's thirty-fifth anniversary in September 2023 his vision was summarized as the aim that machines should understand human beings, so that people would no longer have to communicate with computers in binary code.⁴ IDSIA began as a private institute financed by the Dalle Molle foundation. In 2000 it became a public research institute affiliated with both USI and SUPSI. Since then it has acted as a "bridge" between the two institutions, working "from fundamental to applied research" and transferring its knowledge into applications.⁵ Its offices are on the East Campus of USI and SUPSI in Lugano-Viganello. In 2023 the institute employed 130 people. It was led by a director, Andrea Emilio Rizzoli, and three scientific directors, Jürgen Schmidhuber, Marco Zaffalon and Cesare Alippi.⁶ The institute's research covers machine learning, neural networks, robotics, computer vision, probabilistic reasoning, optimization, computational mathematics, cybersecurity and humancomputer interaction.⁷

Long short-term memory and the history of deep learning

The scientific reputation of IDSIA rests above all on the work of Jürgen Schmidhuber and his group on recurrent neural networks. In 1997 Sepp Hochreiter and Jürgen Schmidhuber published the article "Long Short-Term Memory" in the journal Neural Computation. The article described a recurrent neural network architecture with memory cells and gates that could learn dependencies over long sequences of data, a problem that earlier recurrent networks could not solve because the error signals used in training either vanished or exploded over time.⁸ Hochreiter had developed the analysis of the vanishing gradient problem in his diploma thesis of 1991 at the Technical University of Munich, where Schmidhuber was then his supervisor. The LSTM architecture was refined at IDSIA in the following years by Schmidhuber's doctoral students and collaborators, among them Felix Gers, Alex Graves and others, who added the "forget gate" and developed training methods for sequence labelling.⁹

The practical breakthrough came in the late 2000s. According to Schmidhuber's own account, an LSTM trained with the connectionist temporal classification method of his team became in 2009 the first recurrent neural network to win international pattern recognition competitions, in connected handwriting recognition. In 2011 a deep feedforward network of his team, DanNet, trained on graphics processors, became the first network of its kind to win computer vision contests and achieved what the group described as superhuman performance in a traffic sign recognition competition. In May 2015 the group published the Highway Network, a very deep feedforward network with hundreds of layers, of which the residual network (ResNet) published in December 2015 by researchers at Microsoft is, in Schmidhuber's description, an opengated version.¹⁰ Between 2009 and 2012 the neural networks developed at the institute won eight international competitions in pattern recognition and machine learning.¹¹

LSTM became one of the most widely used neural network architectures of the 2010s. Schmidhuber's page lists its use for speech recognition on Android smartphones, for machine translation by Google Translate and Facebook, for Apple's Siri and QuickType and for the answers of Amazon's Alexa.¹² These claims come from the author of the method. They are consistent with the publications of the companies concerned, which described LSTM-based systems in their speech and translation products in the mid-2010s, before the transformer architecture introduced in 2017 replaced recurrent networks in most language applications.¹³ The attribution of credit in the history of deep learning is contested. Schmidhuber has been a prominent participant in the debate about the priority of ideas. For the purposes of this book it is sufficient to note that some of the methods on which modern machine learning rests were developed in Lugano, and that the institute's work is cited in the standard accounts of the field.¹⁴

Notes

1. Wikipedia (Italian), "Angelo Dalle Molle", https://it.wikipedia.org/wiki/Angelo_Dalle_Molle (consulted as a last resort) (born in Mestre; creator of the artichoke aperitif Cynar in 1952; founder of the Dalle Molle foundation in Lugano for research on artificial intelligence). The article gives 2001 as the year of death, while the institute's own page gives 1908–2002 (below). ↩
2. SUPSI, IDSIA, "Istituto" (above); USI, "Istituto Dalle Molle di studi sull'intelligenza artificiale (IDSIA USI-SUPSI)", https://search.usi.ch/it/unita-organizzative/351. ↩
3. SUPSI, IDSIA, "Istituto" (above) (founded in 1988 in Lugano by Angelo Dalle Molle, 1908–2002). ↩
4. "I 35 anni dell'IDSIA", laRegione, 27 September 2023, https://www.laregione.ch/cantone/ticino/1700339/servizi. ↩
5. SUPSI, IDSIA, "Istituto" (above) (affiliated with USI and SUPSI since 2000); USI, IDSIA unit page (above) ("bridge"; "from fundamental to applied research"). The transformation into a public research institute in 2000 is also stated in Wikipedia, "Dalle Molle Institute for Artificial Intelligence Research", https://en.wikipedia.org/wiki/Dalle_Molle_Institute_for_Artificial_Intelligence_Research (consulted as a last resort). ↩
6. laRegione, 27 September 2023 (above). ↩
7. USI, IDSIA unit page (above), listing more than eighty fields of competence. ↩
8. Sepp Hochreiter and Jürgen Schmidhuber, "Long Short-Term Memory", Neural Computation 9 (1997), no. 8, 1735–1780, as cited on Jürgen Schmidhuber's page on the most cited neural networks, https://people.idsia.ch/~juergen/most-cited-neural-nets.html (reference [LSTM1]). ↩
9. Felix A. Gers, Jürgen Schmidhuber and Fred Cummins, "Learning to Forget: Continual Prediction with LSTM", Neural Computation 12 (2000), no. 10, 2451–2471 (reference [LSTM2] on the same page, above); A. Graves, S. Fernández, F. Gomez and J. Schmidhuber, "Connectionist Temporal Classification: Labelling Unsegmented Sequence Data with Recurrent Neural Networks", ICML 2006, Pittsburgh, as cited in Jürgen Schmidhuber, "Annotated History of Modern AI and Deep Learning" (2022), https://people.idsia.ch/~juergen/deep-learning-history.html, which also describes Hochreiter's diploma thesis of 1991 at TU Munich under Schmidhuber's supervision. The thesis and supervision are also stated in Wikipedia, "Jürgen Schmidhuber", https://en.wikipedia.org/wiki/J%C3%BCrgen_Schmidhuber (consulted as a last resort). ↩
10. Jürgen Schmidhuber, personal home page at IDSIA, https://people.idsia.ch/~juergen/ ("In 2009, the CTC-trained Long Short-Term Memory (LSTM) of his team was the first recurrent NN to win international pattern recognition competitions"; "In 2011, the DanNet of his team was the first feedforward NN to win computer vision contests, achieving superhuman performance"; Highway Net of May 2015 and ResNet of December 2015). ↩
11. Wikipedia, "Dalle Molle Institute for Artificial Intelligence Research" (above) ("won eight international competitions in pattern recognition and machine learning", 2009–2012). ↩
12. Schmidhuber, home page (above). ↩
13. H. Sak, A. Senior, K. Rao, F. Beaufays and J. Schalkwyk (Google Speech Team), "Google voice search: faster and more accurate", Google Research Blog, September 2015, as cited in Schmidhuber, "Annotated History" (above), which also dates the use of LSTM in Google Translate to 2016 and the transformer paper of Ashish Vaswani et al., "Attention Is All You Need", to 2017. ↩
14. See e.g. Jürgen Schmidhuber, "Deep Learning in Neural Networks: An Overview", Neural Networks 61 (2015), 85–117 (reference [DL1] in Schmidhuber, "Annotated History", above), and the "Annotated History" itself, which sets out his view of the priority disputes. ↩

For general information only. This article does not constitute legal or tax advice, does not replace advice on an individual case and creates no lawyer-client relationship. It reflects the legal position as of October 2026.