Detection error in latent spaces also double-descends: high-dimensional theory predicts miss rates at a fixed false alarm
Synopsis
Once foundation models encode images, text and audio into high-dimensional latent vectors, how detection and classification error probabilities vary with the ratio of model size to training samples has lacked a computable characterization. The work derives explicit large-deviation rates for linear and quadratic decision statistics and finds the rate depends non-monotonically on the shape parameter, including double descent and, in some regimes, multiple descent; in latent-space experiments with DINOv3, Llama-Nemotron-Embed-1B-v2 and BEATs, predicted rates and detection probabilities at fixed false alarm track the measured curves closely.
Interpretation
The error decay rate of detection and classification does not worsen monotonically as models grow relative to data in the high-dimensional regime; it can double-descend, and in some settings multiple-descend. Double descent had been characterized mainly for regression, while the impact of high dimensionality on detection and classification error probabilities remained largely unexplored. For the linear decision statistic the rate improves again after the critical threshold; when the population spectral distribution has support points spaced uniformly on a logarithmic scale, the rate shows several local maxima as the regularization parameter varies.
The approach maps observations through foundation-model encoders into a latent space, then applies a Neyman–Pearson decision statistic on that high-dimensional representation. Statistics designed for detection at fixed false alarm cannot be applied directly to real structured data, so the decision is moved to the latent representation. Experiments pair DINOv3 with Airbus and FireRisk, Llama-Nemotron-Embed-1B-v2 with HumAID, and BEATs with SONYC-UST-V2; theoretical rate curves agree with empirical ones overall, with a small vertical shift on FireRisk.
Given a false-alarm level, the theoretical formulas predict the corresponding miss-detection probability, and the approximate log-likelihood ratios beat logistic regression and linear SVM baselines at small false-alarm levels. Standard baselines require fine-tuning, whereas the approximate log-likelihood ratios need essentially none. In ROC comparisons across vision, text and audio, the approximate log-likelihood ratios are systematically better at small false-alarm levels and comparable or better at larger ones.
Perspective
The demonstrated scope is binary detection and multi-class classification in foundation-model latent spaces, with data from DINOv3 on Airbus and FireRisk, Llama-Nemotron-Embed-1B-v2 on HumAID, and BEATs on SONYC-UST-V2, plus an extension to BERT-base and ALBERT-base text representations. The theory assumes a shared covariance across the two hypotheses; the experimental data show covariance mismatch, and the prediction formulas are extended to fit it.
The population spectral distribution and signal direction must be estimated and may be intractable in some settings; degenerate covariance cases fall outside the current analysis. For the quadratic statistic the rate is in some cases only an upper bound, and the large-deviation principle must be verified case by case.
