DI-SCUSUM
A training-free change detection algorithm that constructs a smoothed density estimate by adding Gaussian noise to finite samples and computes a detection statistic using the Hyvarinen score. It requires neither a pretrained model nor knowledge of the distributional forms before and after the change.
Timeline (last 90 days)
DI-SCUSUM (Diffusion-Integrated Score CUSUM) is a completely training-free change detector that constructs a smooth density estimate by adding Gaussian noise to samples and computes the detection statistic with the aid of the Hyvärinen score.
After DI-SCUSUM adds Gaussian noise to limited samples, it can obtain an analytically computable smooth density estimate whose score can be directly derived in the form of the convolution of the samples and the noise, thereby requiring no training of a neural network.
Under the assumption that observations follow a fixed empirical distribution, the average post-change increment of DI-SCUSUM is proportional to the KL divergence from the smoothed post-change distribution to the smoothed pre-change distribution.
The false alarm probability of DI-SCUSUM decays exponentially as the threshold increases.
Under a fixed threshold and incremental scaling, the detection delay of DI-SCUSUM is inversely proportional to the KL divergence.
In calibrated anisotropic Gaussian simulations, DI-SCUSUM performs close to likelihood-ratio CUSUM while reducing detection delay by about 91% compared with score-based CUSUM (SCUSUM).
On the MNIST handwritten digits and Oxford-IIIT Pet datasets, DI-SCUSUM exhibits lower conditional detection delay than SCUSUM at comparable false alarm levels.
DI-SCUSUM can work directly from limited samples, without requiring large amounts of labeled data or training time.
DI-SCUSUM is applicable to scenarios such as equipment fault warning, network anomaly detection, or medical monitoring that require rapid detection of distribution changes with limited data.
All Facts (9)
DI-SCUSUM (Diffusion-Integrated Score CUSUM) is a completely training-free change detector that constructs a smooth density estimate by adding Gaussian noise to samples and computes the detection statistic with the aid of the Hyvärinen score.
50%UnverifiedAfter DI-SCUSUM adds Gaussian noise to limited samples, it can obtain an analytically computable smooth density estimate whose score can be directly derived in the form of the convolution of the samples and the noise, thereby requiring no training of a neural network.
50%UnverifiedUnder the assumption that observations follow a fixed empirical distribution, the average post-change increment of DI-SCUSUM is proportional to the KL divergence from the smoothed post-change distribution to the smoothed pre-change distribution.
50%UnverifiedThe false alarm probability of DI-SCUSUM decays exponentially as the threshold increases.
50%UnverifiedUnder a fixed threshold and incremental scaling, the detection delay of DI-SCUSUM is inversely proportional to the KL divergence.
50%UnverifiedIn calibrated anisotropic Gaussian simulations, DI-SCUSUM performs close to likelihood-ratio CUSUM while reducing detection delay by about 91% compared with score-based CUSUM (SCUSUM).
50%UnverifiedOn the MNIST handwritten digits and Oxford-IIIT Pet datasets, DI-SCUSUM exhibits lower conditional detection delay than SCUSUM at comparable false alarm levels.
50%UnverifiedDI-SCUSUM can work directly from limited samples, without requiring large amounts of labeled data or training time.
50%UnverifiedDI-SCUSUM is applicable to scenarios such as equipment fault warning, network anomaly detection, or medical monitoring that require rapid detection of distribution changes with limited data.
50%