Feedback Loop Collapse

Feedback loop collapse refers to the degradation of a machine learning model's performance due to training on its own outputs, leading to a self-reinforcing cycle of error propagation. This phenomenon typically occurs in generative models where the model generates data that is subsequently used for further training. Mathematically, this can be represented by the recursive relationship between the model's output and its input, where the quality of the generated data diminishes over iterations. The collapse can be quantified using metrics such as the expected log-likelihood of the generated data, which may decrease as the model becomes increasingly reliant on its own potentially flawed outputs. Feedback loop collapse is a significant concern in applications involving self-supervised learning and is related to broader issues of model stability and reliability in machine learning.