09. August 2021 : A novel approach for training artificial neural networks

Clay V, König P, Kühnberger KU and Pipa G (2021).
Fast concept mapping: The emergence of human abilities in artificial neural networks when learning embodied and self-supervised.
https://arxiv.org/abs/2102.02153 arXiv:2102.02153

Science has proved that humans perceive the world and gradually understand and learn about its underlying concepts by self-supervised interaction and embodiment. By contrast, most artificial neural networks used for object detection and recognition are trained in a fully supervised setup requiring large data sets of labeled examples. In our study, we built a simulated world, which an artificial agent was left to explore and interact with self-supervised and curiosity-driven. Subsequently, the representations learned were used to associate semantic concepts. This "fast concept mapping" method uses correlated firing patterns of neurons to define and detect semantic concepts. Strikingly, our method already identified objects with as little as one labeled example. This training method promises to be a feasible and resources saving strategy for artificial neural networks to learn meaningful representations.