
bioRxiv (Cold Spring Harbor Laboratory), Год журнала: 2024, Номер unknown
Опубликована: Май 22, 2024
Abstract We show that neural networks can implement reward-seeking behavior using only local predictive updates and internal noise. These are capable of autonomous interaction with an environment switch between explore exploit behavior, which we is governed by attractor dynamics. Networks adapt to changes in their architectures, environments, or motor interfaces without any external control signals. When have a choice different tasks, they form preferences depend on patterns noise initialization, these be biased network architectures changing learning rates. Our algorithm presents flexible, biologically plausible way interacting environments requiring explicit environmental reward function, allowing for both highly adaptable autonomous. Code available at https://github.com/ccli3896/PaN .
Язык: Английский