Technische Universität München and the 1997 LSTM Paper

TUM's AI history includes Sepp Hochreiter's contribution to the 1997 LSTM paper, which introduced gated memory cells for recurrent-network learning.

kindUniversity
founded1868
sitetum.de

Updated

A technical university with a recurrent-network contribution

Technische Universität München, known in English as the Technical University of Munich or TUM, traces its origins to the New Polytechnic School founded in Munich in 1868. The historical names describe the university's development, rather than separate present-day institutions.

In the 1997 Long Short-Term Memory paper, Sepp Hochreiter was affiliated with TUM and co-author Jürgen Schmidhuber with IDSIA in Lugano. They introduced recurrent-network memory cells with controlled error flow and multiplicative gates, designed to learn dependencies across long time intervals. Their reported tests used artificial sequence tasks, including delays exceeding a thousand steps. The reported results concern those specified artificial tasks, rather than every later sequence-modeling application.

Sources

  1. [1]
    History of Technical University of Munich (TUM)

    Technical University of Munich · Apr 25, 2012

  2. [2]
    Long Short-Term Memory

    Sepp Hochreiter and Jürgen Schmidhuber · 1997

Technische Universität München history

1997

■
LSTM gives networks a memory

With IDSIA

Hochreiter and Schmidhuber solved the vanishing gradient problem with a gated cell that could hold information across long sequences. It ran sequence modeling for two decades, until attention took the job.

Neural Computation
Research