Jürgen Schmidhuber is a German computer scientist and co-scientific director of the Dalle Molle Institute for Artificial Intelligence Research (IDSIA) in Lugano, Switzerland. He is best known for co-inventing the long short-term memory (LSTM) network with his then-student Sepp Hochreiter in 1997, one of the most widely used Recurrent neural network architectures for two decades of Natural language processing and sequence modeling work before the rise of the Transformer (architecture).
Career and research
Born in 1963 in Munich, Schmidhuber earned his PhD from the Technical University of Munich in 1991 and has spent most of his career at IDSIA, where he built one of the influential European labs for Deep learning research well before the field's mainstream 2012 breakout. IDSIA's group also produced early practical successes in handwriting and image recognition competitions in the late 2000s using GPU-accelerated deep networks, work that predated but paralleled the more widely publicized AlexNet result of 2012. Beyond LSTM, his lab produced early work on recurrent network training, artificial curiosity and intrinsic motivation for reinforcement learning agents, and Highway Networks, a precursor to the residual connections later used in very deep image classifiers.
Priority disputes
Schmidhuber has become almost as well known for publicly and repeatedly arguing that the field under-credits his and his collaborators' work relative to researchers such as Geoffrey Hinton, Yann LeCun and Yoshua Bengio. He has claimed that key ideas behind Backpropagation, the Generative adversarial network (crediting an earlier adversarial curiosity formulation), and elements of attention mechanisms used in transformers were anticipated by earlier IDSIA-affiliated work that he says was overlooked when Hinton, LeCun and Bengio received the 2018 Turing Award. These disputes, aired in blog posts, conference remarks and interviews, have made him a controversial but closely followed figure within the Artificial intelligence research community, where questions of precedence in a fast-moving field are frequently contested.
Later ventures
Schmidhuber co-founded NNAISENSE, a company applying recurrent and reinforcement-learning based methods to industrial and robotics problems, and has continued to publish on general-purpose learning systems and the theoretical limits of intelligence, maintaining that many ideas now central to Machine learning and modern AI agent research trace back further than commonly acknowledged. He has also written extensively about formal theories of optimal, resource-bounded universal problem solvers, work that remains largely outside the mainstream of industrial deep learning but that he presents as a more principled long-term path toward general-purpose artificial intelligence than incremental scaling alone.