# Jürgen Schmidhuber > 1963– · Computer Scientist, AI Researcher > > **Recorded contribution:** LSTM (with Hochreiter); RNNs; formal theory of curiosity/creativity; compressed network search ## How to use this dossier Read for a causal chain, not a hero story: inherited problem → contribution → mechanism → downstream capability → limit. Then close the page and complete the reconstruction exercise from memory. ## 1. Historical orientation Jürgen Schmidhuber supervised Sepp Hochreiter’s work on the vanishing-gradient problem and co-authored the 1997 Long Short-Term Memory paper with him; he also researched recurrent networks, compression, curiosity, and algorithmic information. LSTM supplied a trainable memory mechanism for sequences before attention-based architectures became dominant. This work belongs to the history of making machine behavior depend on representations, evidence, objectives, and evaluation rather than on a separate hand-written rule for every case. The chronology is used causally: it connects the inherited constraint to an implementable mechanism and then to later reuse, instead of treating fame, job title, or eventual market success as the explanation. ## 2. The problem inherited Ordinary recurrent networks repeatedly multiplied gradients through time, causing long-range learning signals to vanish or explode before reaching early events. The first-principles difficulty is not simply “make a machine intelligent”: it is to specify what is represented, where evidence comes from, how a procedure changes with evidence, and what observation would count as failure. ## 3. The central contribution An LSTM cell uses a persistent state and learned gates that control writing, forgetting, and exposing information, creating a path along which gradients can remain more stable. Its importance therefore lies in an inspectable learning or search mechanism, not in an anthropomorphic claim about the system understanding as a person does. ## 4. Reconstruct the mechanism 1. Compute candidate content from the current input and previous hidden state. State the task, representation, and success measure before selecting an algorithm. 2. Use input and forget gates to decide what enters and remains in the cell state. Trace where evidence or feedback changes internal state; do not hide learning behind a product label. 3. Use an output gate to expose selected state as the next hidden representation. Run the resulting procedure on a small case where every intermediate value can be inspected. 4. Unroll through a long sequence and inspect gradient magnitude, saturation, memory capacity, and failure on extrapolation. Change the data, objective, or environment and locate the first place behavior ceases to generalize. ## 5. What changed downstream - LSTM became central to speech recognition, handwriting, translation, and sequence modeling and remains an important conceptual ancestor of gated memory systems. - Downstream systems inherited both a reusable method and a warning: benchmark performance depends on the data-generating process and evaluation contract. - The transferable first-principles lesson is to separate the artifact named in “LSTM (with Hochreiter); RNNs; formal theory of curiosity/creativity; compressed network search” from the mechanism, surrounding institution, and evidence that allowed later systems to depend on it. ## 6. Attribution, limits, and uncertainty - LSTM is jointly credited to Hochreiter and Schmidhuber and was enabled by prior recurrent-network and backpropagation research. Later implementations and adoption involved many teams. Priority disputes around deep learning require paper-specific evidence, not broad “father” labels. - Later success does not retroactively prove that every historical motivation, cognitive analogy, or priority claim was correct. - The subject is living or the registry has no death year; current titles and institutional affiliations are treated as dated snapshots verified on 2026-08-09, not permanent identity claims. ## 7. Reconstruction lab Implement a minimal LSTM and simple recurrent baseline on a delayed-bit task. Plot gradient norms by time step and vary sequence length until each model fails. Report the representation, objective, update/search rule, held-out test, and one deliberately adversarial example. ## 8. Evidence trail - [Long Short-Term Memory](https://www.bioinf.jku.at/publications/older/2604.pdf) — Neural Computation - [Jürgen Schmidhuber](https://en.wikipedia.org/wiki/J%C3%BCrgen_Schmidhuber) — Wikipedia contributors · overview and bibliography - [Jürgen Schmidhuber structured identity record](https://www.wikidata.org/wiki/Q92735) — Wikidata contributors · CC0 --- *Research checked 2026-08-09. Dates, roles, and claims about living people are historical snapshots. Linked sources remain the authority; this dossier is original instructional synthesis.*