Schmidhuber slams NeurIPS for honoring GAN paper he says republished his 1990 Artificial Curiosity work

SchmidhuberAI · x · 2026-09-04

Jürgen Schmidhuber criticized the GAN paper winning the NeurIPS 2024 Test of Time Award, claiming it republished his 1990 Artificial Curiosity principle under a new name without citation. He notes the 1990-91 papers already described two neural networks fighting each other: a generative model with adaptive stochastic Gaussian units producing data, and a predictor learning by gradient descent — published when compute was 100,000x more expensive than in 2014.

Booster Wirakumon argued the history of deep learning is being written wrong: while recent engineering advances matter, the fundamental contributions from 30+ years ago deserve credit, and the GAN authors at minimum should have acknowledged the prior work — which they never did.

Original post →

More from Fun

Fun channel →