Recovering Data Drift Regimes from Neural Network Weights

This paper explores whether discrete temporal regimes within drifting data streams can be identified by analyzing the weights of neural networks trained on that data. The researchers propose fitting a Hidden Markov Model (HMM) to the chronologically ordered trajectory of model weights, which are trained on consecutive temporal windows of the data. This approach aims to recover latent states that partition the data's timeline into coherent phases. The study was conducted across two domains known for temporal drift: multimodal misinformation detection (Fakeddit dataset) and sentiment analysis (Yelp dataset). Results show that classifiers generalize better to data from windows sharing the same recovered latent state than to data across state boundaries. These states, estimated solely from model weights, surprisingly correlate more strongly with shifts in the data's class distribution than with the weight-space geometry used for estimation, indicating they capture relevant structure beyond mere temporal proximity.