Search papers, labs, and topics across Lattice.
This paper addresses the challenge of detecting distributional shifts in sequential data where neither the pre- nor post-change distributions have a closed form. By leveraging a conditional diffusion model and mapping pre-change observations to standard Gaussian latent variables, the authors derive a test statistic based on Maximum Mean Discrepancy, allowing for effective online change detection. The proposed method successfully identifies various types of distributional shifts, including covariance rotations, without relying on parametric assumptions, showcasing its robustness in practical applications.
Detecting arbitrary distributional shifts in sequential data without parametric assumptions could revolutionize how we approach change detection in complex systems.
A rapidly growing range of sequential data tasks, such as identifying trend reversals in financial markets, auto-segmenting video and audio recordings, detecting changes in movement direction from motion sensors cannot be fully addressed without detection of distributional shifts in time-ordered data. We consider a sequential change-point detection problem where the conditional density switches at an unknown time, yet neither the pre- nor post-change distribution admits a closed-form. Classical likelihood-ratio statistics are inapplicable in this settings. A conditional diffusion model, trained on pre-change-point data with a frozen context encoder, defines a deterministic bijection via the probability flow ODE. Pre-change observations are mapped onto standard Gaussian latent variables. Post-change observations, processed through the same frozen map, deviate from this reference. We employ the Maximum Mean Discrepancy as the test statistic, derive closed-form expressions for its components under the Gaussian null, and establish its asymptotic distribution as a degenerate U-statistic. Afterwards we apply an online detection procedure of Shiryaev--Roberts to the resulting statistic with exact threshold calibration. The method detects arbitrary distributional shifts, including covariance rotations and higher-order structural breaks, without parametric assumptions on either regime.
