[Paper Review] Sequential Drift Detection in Deep Learning Classifiers

Samuel Ackerman, Parijat Dube+1 more|arXiv (Cornell University)|Jul 31, 2020

Data Stream Mining Techniques10 references5 citations

TL;DR

This paper proposes a sequential drift detection framework for deep learning classifiers using neural network embeddings and the Cramér-Price-Mann (CPM) sequential test to control false alarm rates. By formulating drift detection as a sequential hypothesis test on univariate embedding divergence, the method achieves superior false alarm control compared to non-sequential Mann-Whitney testing, with only a modest increase in detection delay, as validated by a novel loss function balancing delay and error rates across multiple datasets.

ABSTRACT

We utilize neural network embeddings to detect data drift by formulating the drift detection within an appropriate sequential decision framework. This enables control of the false alarm rate although the statistical tests are repeatedly applied. Since change detection algorithms naturally face a tradeoff between avoiding false alarms and quick correct detection, we introduce a loss function which evaluates an algorithm's ability to balance these two concerns, and we use it in a series of experiments.

Motivation & Objective

• To address the challenge of controlling false alarm rates in data drift detection for deployed deep learning models.
• To improve upon non-sequential drift detection methods like Mann-Whitney testing, which suffer from high false alarm rates due to repeated testing.
• To develop a sequential statistical framework that maintains a known, low Type-I error rate (e.g., α = 0.05) over time.
• To introduce a customizable loss function that quantitatively evaluates drift detection performance based on detection delay and false alarm rate.
• To empirically compare sequential detection using CPM with non-sequential Mann-Whitney testing across diverse data drift scenarios.

Proposed method

• The method uses embeddings from the penultimate layer of a trained deep neural network as a univariate representation of input data distribution shifts.
• It computes divergence between the current embedding distribution and a baseline using metrics like cosine distance or Kullback-Leibler divergence.
• The Cramér-Price-Mann (CPM) sequential change detection algorithm is applied to monitor the divergence over time, enabling statistical control of the false alarm rate.
• The framework treats drift detection as a sequential hypothesis test, where the probability of a false alarm is bounded by a user-defined significance level α.
• A novel loss function is introduced that combines detection delay and false alarm rate, with adjustable weights to reflect user priorities.
• The method is evaluated across 15 different drift scenarios involving class prior probability shifts, using both linear and step-wise contamination patterns.

Experimental results

Research questions

• RQ1 Can sequential change detection methods like CPM reduce false alarm rates in deep learning drift detection compared to non-sequential approaches such as repeated Mann-Whitney testing?
• RQ2 How does the trade-off between detection delay and false alarm rate vary across different drift types (linear vs. step) and data distributions?
• RQ3 To what extent can a custom loss function effectively balance detection speed and false alarm control in drift detection algorithms?
• RQ4 Does the use of deep network embeddings enable robust, univariate representation of complex data drift in high-dimensional image inputs?
• RQ5 In what scenarios does the sequential approach outperform the non-sequential Mann-Whitney method in terms of overall detection performance?

Key findings

• The sequential CPM-based method achieved zero false alarms across all 15 drift scenarios, while the non-sequential Mann-Whitney test had an average false alarm rate of 0.089 (KLD) and 0.315 (cosine distance).
• The Mann-Whitney method had a lower average detection delay (14.267 vs. 19.6 for KLD; 12.6 vs. 18.067 for cosine) but at the cost of significantly higher false alarm rates.
• The overall loss function score was consistently lower (better) for the sequential method: -0.451 (KLD) and -0.426 (cosine) compared to -0.527 and -1.366 for Mann-Whitney, respectively.
• In rare cases where the false alarm penalty was heavily downweighted (very low C1), the Mann-Whitney method slightly outperformed the sequential method, but this was limited to small parameter regions.
• The sequential method maintained strong performance across diverse drift types, including both gradual (linear) and abrupt (step) contamination patterns.
• The use of neural network embeddings successfully reduced the high-dimensional image drift problem to a univariate statistical monitoring task, enabling effective sequential analysis.