Jürgen Schmidhuber Proposes Artificial Curiosity

This pioneering idea in machine learning during the supposed “AI winter” was a proposal of a generative adversarial network (GAN), which would not be formally realized until 2014. This was an example of an important development in the creative application of neural networks, demonstrating how far-reaching the ideas of the past decade-plus extend. The framework involved two competing neural networks: a generator (controller) that creates outputs and a predictor (world model) that anticipates the consequences of these outputs. The networks engage in a minimax game, where the generator aims to maximize the predictor’s error, while the predictor minimizes it. This early concept laid the groundwork for future developments in generative models, including the now widely used Generative Adversarial Networks (GANs).