The implicit regularizing effect of stochastic resetting in deep learning analysis of anomalous diffusion

Here, we investigate whether stochastic resetting, a technique that periodically reverts training to beneficial checkpoints, recently explored in the context of deep learning with noisy labels, can improve the decoding of anomalous diffusion trajectories, a task made challenging by noise and limited trajectory lengths, which often render subtle differences between diffusion processes indistinguishable. First, we find that incorporating stochastic resets of neural network parameters improves the validation loss across different hyperparameters and noise levels, confirming its applicability in trajectory decoding tasks. Second, we find that the relative benefit of resetting increases with trajectory length, and we offer a mechanistic explanation supported by minibatch-gradient ensemble diagnostics that links this observation to the underlying optimization dynamics. Furthermore, we observe that there exists an optimal resetting probability that yields the best performance, highlighting the importance of tuning this hyperparameter. Building on this insight, we introduce time-varying resetting mechanisms that dynamically adjust the resetting probability during training. Our results show that these mechanisms often match or surpass the performance of the best fixed-resetting probabilities and offer a solid basis for designing effective dynamic resetting strategies for regularization.