[Paper Review] Novelty detection using ensembles with regularized disagreement.

Alexandru Ţifrea, Eric Stavarache+1 more|arXiv (Cornell University)|Dec 10, 2020

Anomaly Detection Techniques and Applications54 references4 citations

TL;DR

This paper proposes a novel ensemble-based method for out-of-distribution (OOD) detection that leverages regularized disagreement to improve detection of novel classes, even when OOD data is not available during training. It significantly outperforms state-of-the-art methods on standard benchmarks like SVHN, CIFAR-10, CIFAR-100, and medical image datasets, particularly in detecting unseen classes.

ABSTRACT

Despite their excellent performance on in-distribution (ID) data, machine learning-based prediction systems often predict out-of-distribution (OOD) samples incorrectly while indicating high confidence. Instead, they should flag samples that are not similar to the training data, for example, when new classes emerge over time. Even though current OOD detection algorithms can successfully distinguish completely different data sets, they fail to reliably identify samples from novel classes. We develop a new ensemble-based procedure that promotes model diversity and exploits regularization to limit disagreement to only OOD samples, using a batch containing an unknown mixture of ID and OOD data. We show that our procedure significantly outperforms state-of-the-art methods, including those that have access, during training, to data that is known to be OOD. We run extensive comparisons of our approach on a variety of novel-class detection scenarios, on standard image data sets such as SVHN/CIFAR-10/CIFAR-100, as well as on new disease detection on medical image data sets.

Motivation & Objective

• To address the limitation of deep learning models in detecting out-of-distribution (OOD) samples, especially novel classes not seen during training.
• To improve OOD detection performance when no known OOD data is available during model training.
• To develop a method that promotes model diversity while restricting disagreement to only OOD samples through regularization.
• To evaluate the method on diverse scenarios, including standard image datasets and real-world medical image classification tasks.

Proposed method

• The method uses an ensemble of deep neural networks to increase model diversity and improve robustness to OOD inputs.
• It introduces a regularized disagreement loss that limits model disagreement to only OOD samples, encouraging agreement on in-distribution (ID) data.
• The training process uses a mixed batch containing both ID and OOD samples, enabling the model to learn to distinguish OOD inputs through disagreement patterns.
• The regularization component ensures that disagreement is suppressed on ID data, making the model more confident on known distributions.
• The approach does not require access to known OOD data during training, unlike many existing methods.
• The final prediction is based on ensemble outputs, with high disagreement used as a signal for OOD detection.

Experimental results

Research questions

• RQ1 Can an ensemble-based method detect novel classes without requiring access to known OOD data during training?
• RQ2 How does regularized disagreement improve OOD detection performance compared to standard ensemble methods?
• RQ3 Does the method generalize across diverse data distributions, including medical imaging and standard vision benchmarks?
• RQ4 How does the method perform in scenarios where OOD samples are from novel classes rather than completely different datasets?
• RQ5 Can the method achieve state-of-the-art performance on standard image datasets like CIFAR-10, CIFAR-100, and SVHN?

Key findings

• The proposed method significantly outperforms state-of-the-art OOD detection methods, even those that have access to known OOD data during training.
• The method achieves strong performance on CIFAR-10, CIFAR-100, and SVHN, demonstrating robustness across standard vision benchmarks.
• It effectively detects novel classes in medical image datasets, showing applicability beyond standard image recognition tasks.
• The model maintains high confidence on in-distribution data while generating high disagreement on OOD inputs, indicating reliable detection.
• The regularization of disagreement leads to improved detection performance by suppressing false positives on ID data.
• The method is effective even when OOD data is not explicitly provided during training, making it practical for real-world deployment.