Technical University of Munich

What TUM is

The Technical University of Munich is a public technical university in Munich, in the German state of Bavaria. It was founded on 12 April 1868 by King Ludwig II as the Polytechnische Schule München ^[1]. Its home page presents the university as "The Entrepreneurial University" and refers to a staff of 10,000 ^[5].

TUM's home page lists seven schools: Computation, Information and Technology; Engineering and Design; Natural Sciences; Life Sciences; Medicine and Health; Management; and Social Sciences and Technology ^[5]. The same page names six research goals, among them Digitalization, AI and Quantum Technologies and Foundations of Life ^[5]. Beyond the schools, its integrative research centres include the Munich Data Science Institute and the Munich School of Robotics and Machine Intelligence ^[1]. Industry ties are part of how the university describes itself: its home page lists 27 companies and organizations as Partners of Excellence, including Airbus, BMW, Siemens and SAP ^[5].

How computing grew at TUM

Computer science did not start out as a department of its own. It was taught within the Department of Mathematics from 1967, and became an independent Department of Informatics only in 1992 ^[1]. That timing matters for the university's later reputation in machine learning, because the work most closely associated with TUM in this field dates from the years when computer science was still taught inside the mathematics department.

Two names in sequence learning

Sepp Hochreiter's 1991 diploma thesis at TUM analysed why error signals vanish or explode when recurrent networks are trained by gradient descent, and it contained the first LSTM results; he completed his doctorate at the same university in 1999 under Wilfried Brauer ^[2].

The problem the thesis addressed is easy to state. A recurrent network processes a sequence by feeding information from one step back into the next, and it learns by gradient descent — adjusting its parameters a little in the direction that reduces error. When such a network is unrolled over many steps, the error signal that has to travel back through them can shrink towards nothing or grow without bound. If it shrinks, the early parts of the sequence receive almost no learning signal; if it grows, training becomes unstable. The LSTM design that came out of this work introduced a memory cell with gates that control what is written into the cell, kept, and read out, so that information can be carried across many steps instead of decaying.

The 1997 paper "Long Short-Term Memory", published in Neural Computation, lists Hochreiter at the Technische Universität München and Jürgen Schmidhuber at IDSIA ^[3].

Schmidhuber's own connection to TUM is long. He studied computer science there, earning his diploma in 1987, his doctorate in 1991 and his habilitation in 1993 ^[4]. He later taught at TUM from 2004 to 2009, where he headed the Cog Bot Lab ^[4].

Place in the field

The link between a nineteenth-century polytechnic and modern sequence models is concrete rather than rhetorical: the paper that introduced long short-term memory carries a Munich affiliation on its author list ^[3], and both of its authors passed through the same institution as students or staff ^[2][4]. For readers following how AI research spread across institutions, TUM is a useful case — a European technical university whose computing department grew out of mathematics, and where a diploma thesis on a training difficulty became the starting point for an architecture that later travelled far beyond its original department ^[1][2].