Abstract
Deep Neural Networks (DNNs) are notoriously overconfident when encountering data they were not trained on, a phenomenon known as Out-of-Distribution (OOD) overconfidence. This paper argues that the root cause of this flaw is model over-complexity—specifically, excessive parameter capacity and high nonlinearity. We propose a model simplification framework that utilizes knowledge distillation to reduce capacity and selective ReLU removal to decrease nonlinearity. Our results demonstrate that these simplified models significantly mitigate OOD overconfidence and improve detection performance across standard benchmarks without sacrificing accuracy on the original task.
Introduction
In controlled laboratory settings, DNNs achieve state-of-the-art performance. However, in real-world applications such as autonomous driving or medical diagnosis, they face the dangerous challenge of OOD overconfidence. When presented with unfamiliar inputs, these models often produce incorrect predictions with near-total confidence.
Existing solutions typically rely on post-hoc calibration or training with auxiliary OOD datasets. We argue these are merely symptomatic treatments. Instead, we identify over-parameterization and high nonlinearity as the structural drivers that allow models to "hallucinate" patterns in noise. By simplifying the model architecture, we can create smoother decision boundaries that naturally reflect uncertainty for OOD data.
Methodology
We propose a two-pronged strategy for model simplification:
- Capacity Reduction via Knowledge Distillation (KD): Rather than deploying massive "teacher" networks, we use KD to train smaller "student" networks. The student captures essential features from the teacher while lacking the excess parameters that fuel overconfident OOD predictions.
- Nonlinearity Reduction via Selective ReLU Removal: ReLU activations are the primary source of nonlinearity in DNNs. We introduce the Negative Response Ratio (NRO) to identify ReLU layers with minimal contribution to task performance. These layers are selectively replaced with identity mappings during the distillation process.
Key Insights and Results
The "Goldilocks" Capacity
Our experiments reveal a sweet spot for model size. As shown in the comparison between a ResNet-152 teacher and its distilled students, reducing capacity improves OOD metrics like FPR95 and AUROC up to a point. However, if the model becomes too small (e.g., ResNet-18 in some contexts), it may begin to lose the ability to handle the original In-Distribution (ID) task effectively.
Shifting the Confidence Distribution
Simplification directly impacts the model's confidence on OOD samples. On the CIFAR-100 benchmark, a standard ResNet-50 exhibited a Mean Maximum Confidence (MMC) of 72.12% on OOD data, whereas our simplified ResNet-18 student reduced this to 48.35%.
Synergy and Safety
Unlike many state-of-the-art methods, our approach is data-independent and does not require exposure to outlier data during training. Furthermore, simplification acts as a foundation that enhances existing OOD detection tools (e.g., Energy scores or ReAct). When combined, simplified models outperform specialized mitigation methods like ACET and CEDA, particularly in challenging large-scale scenarios like ImageNet-1k.
Conclusion
Model simplification offers a robust, principled approach to the OOD overconfidence problem. By smoothing the model's response through reduced capacity and nonlinearity, we achieve safer, more reliable predictions. Future work will explore the application of these principles to modern architectures like Vision Transformers (ViTs) and further investigate the theoretical underpinnings of why simplification specifically targets OOD overconfidence.
Summary updated