IDSIA (Dalle Molle Institute for Artificial Intelligence)

What IDSIA is

IDSIA — the Dalle Molle Institute for Artificial Intelligence — is a research institute in the Lugano district of Ticino, in southern Switzerland, at Via la Santa 1, CH-6900 Lugano.^[1][6] It is a joint institute of the Faculty of Informatics of the Università della Svizzera italiana (USI) and the Department of Innovative Technologies of SUPSI, the two institutions that share it.^[6] Andrea Emilio Rizzoli is listed as its director.^[1]

The institute lists its research areas as autonomous robotics, intelligent control for networked systems, geometric and visual computing, machine learning and artificial neural networks, natural language processing and information retrieval, responsible AI and society, and theory and algorithms.^[6] The list is broad, but the pieces fit together: methods for learning and for control, the systems those methods run on, and questions about how such systems should be used.

Origins

IDSIA was founded in 1988 by Angelo Dalle Molle, through the private Fondation Dalle Molle.^[1] It began, in other words, as a privately funded initiative rather than a university department. In 2000 it became a public research institute affiliated with USI and SUPSI.^[1] That change put the institute inside the Swiss public university system while leaving it with its own research agenda; its later status as a joint institute of the two institutions follows from the same arrangement.^[6]

Leadership and machine learning

Jürgen Schmidhuber has directed research at IDSIA since 1995.^[2][5] The institute became known above all for work on artificial neural networks, and the record is concrete: between 2009 and 2012, neural networks developed at the institute won eight international competitions in pattern recognition and machine learning.^[1]

The paper most closely associated with IDSIA is "Long Short-Term Memory", published in Neural Computation in 1997 and written by Sepp Hochreiter of the Technical University of Munich and Schmidhuber of IDSIA.^[3] A recurrent neural network reads a sequence one step at a time and carries an internal state from step to step, so that what it has already seen can influence what it does next. Such networks had long been limited by a practical problem: information from early in a sequence tends to fade as more of the sequence is processed. The LSTM design was aimed at that problem, allowing information to be carried across many steps instead of being lost along the way.

The same theme — making networks hold on to information, or grow deeper — appears again in the 2015 paper "Highway Networks" by Rupesh Srivastava, Klaus Greff and Schmidhuber, which introduced very deep feedforward networks with gated shortcut connections.^[4] A shortcut connection lets information bypass layers rather than passing through every one of them in turn; gating decides how much passes. That is what makes stacking many layers trainable in practice.

Robotics and recognition

A robotics lab focused on intelligent and learning robots, especially swarm and humanoid robotics, was set up at the institute in 2007.^[1] It sits alongside the learning and control work rather than apart from it: robots that learn are both a test case for the algorithms and a research topic in their own right.

Outside recognition came early. A 1997 Business Week survey listed IDSIA among the world's top ten AI laboratories and among the top four in biologically inspired AI.^[1]

Where it sits

The institute's story is a fairly unusual one for an AI lab: a private foundation created by a single founder, folded into public higher education, and then associated with some of the most widely used ideas in neural networks. LSTM and Highway Networks both come out of the same line of work — keeping information available across many steps, or letting it skip across layers — and both carry the IDSIA affiliation in their author lines.^[3][4] The eight competition wins between 2009 and 2012 belong to the same period of neural-network research.^[1] Meanwhile the institute's stated scope is wider than machine learning alone, running from autonomous robotics and intelligent control to natural language processing and questions about responsible AI and society.^[6]