Recovering Temporal Data Drift Regimes from Neural Network Weights

A hidden Markov model (HMM) applied to the weight trajectories of models trained on drifting data streams can successfully recover discrete latent states. Testing on Fakeddit and Yelp datasets shows that classifiers generalize more effectively to data within the same recovered temporal state than across state boundaries.