Out-of-Distribution Detection via Deep Multi-Comprehension Ensemble

TL;DR

The experimental results demonstrate the superior performance of the MC Ensemble strategy in OOD detection compared to both the naive Deep Ensemble method and a standalone model of comparable size, underscoring the effectiveness of the proposed approach in enhancing the model's capability to detect instances outside its training distribution.

Abstract

Recent research underscores the pivotal role of the Out-of-Distribution (OOD) feature representation field scale in determining the efficacy of models in OOD detection. Consequently, the adoption of model ensembles has emerged as a prominent strategy to augment this feature representation field, capitalizing on anticipated model diversity. However, our introduction of novel qualitative and quantitative model ensemble evaluation methods, specifically Loss Basin/Barrier Visualization and the Self-Coupling Index, reveals a critical drawback in existing ensemble methods. We find that these methods incorporate weights that are affine-transformable, exhibiting limited variability and thus failing to achieve the desired diversity in feature representation. To address this limitation, we elevate the dimensions of traditional model ensembles, incorporating various factors such as different weight initializations, data holdout, etc., into distinct supervision tasks. This innovative approach, termed Multi-Comprehension (MC) Ensemble, leverages diverse training tasks to generate distinct comprehensions of the data and labels, thereby extending the feature representation field. Our experimental results demonstrate the superior performance of the MC Ensemble strategy in OOD detection compared to both the naive Deep Ensemble method and a standalone model of comparable size. This underscores the effectiveness of our proposed approach in enhancing the model's capability to detect instances outside its training distribution.

Figures (3)

• Figure 1: (a) Models trained by different initialization ($\Theta_1$, $\Theta_2$) but with the same cross-entropy classification task (CE-SGD) can fall into the same or symmetric loss basin, which can be affine-transformed into the same basin (Re-Basinainsworth2023git). This indicates the two models provide little variability. (b) By contrast, a different comprehension task (SimCLR-SGD) directs the model parameters in other directions. When SimCLR-SGD weights are relocated to the same loss landscape of CE-SGD weights, we can observe the loss barrier between two sets of weights is high so that Re-Basin is not possible, thus increasing the model and feature diversity.
• Figure 2: Two models trained from different hypotheses. When there is a large loss barrier between models, the coupling matrix of features tends to perform more stochastic. The models' architecture is ResNet-18 he2016deep. H1: Initialization $\Theta_1$, Cross-Entropy loss, whole training set. H2: Initialization $\Theta_2$, SimCLR Loss, whole training set. H3: Initialization $\Theta_2$, Cross-Entropy Loss, whole training set. H4: Initialization $\Theta_1$, Cross-Entropy Loss, 80% training set. H5: Initialization $\Theta_2$, Cross-Entropy Loss, another 80% training set.
• Figure 3: t-SNE visualization of penultimate layer features of SupCE, SupCon, SimCLR, MC Ensemble(average), while class 0-9 are ID classes (CIFAR10) and class 10 is OOD (CIFAR100).

Theorems & Definitions (4)

• Theorem 1.1
• proof
• proof
• proof