KLAS: Using Similarity to Stitch Neural Networks for Improved Accuracy-Efficiency Tradeoffs

KLAS is a new framework that automates the selection of neural network stitching configurations by using KL divergence to measure similarity between pretrained models, enabling better accuracy-efficiency tradeoffs. The approach improves upon existing heuristic-based methods and achieves up to 1.21% higher accuracy on ImageNet-1K at equivalent computational cost, or reduces computational requirements by 1.33x while maintaining performance.

KLAS addresses a fundamental challenge in machine learning deployment: how to efficiently select and combine pretrained models for diverse computational environments. The research tackles the combinatorial explosion problem inherent in model stitching, where selecting optimal configurations from O(k²n²) possibilities becomes computationally intractable as model families grow. By grounding stitch selection in a principled similarity metric—KL divergence between intermediate representations—rather than heuristics, the framework provides a generalizable approach that works across different model architectures and families.

The significance of this work lies in its practical implications for edge computing and resource-constrained deployment scenarios. As organizations increasingly need to deploy models across heterogeneous hardware—from mobile devices to cloud servers—the ability to automatically generate optimal accuracy-efficiency tradeoff curves becomes valuable. Current methods rely on manual configuration or simple rules that often miss superior combinations.

KLAS demonstrates measurable improvements: achieving either 1.21% accuracy gains at matched computational budgets or 1.33x efficiency improvements while preserving accuracy. These gains compound when scaled across millions of deployed models. For developers and enterprises managing model deployment pipelines, this translates to better resource utilization and potentially reduced inference costs without sacrificing performance.

The framework's generalizability across model families is particularly noteworthy, suggesting the similarity-based approach captures fundamental principles about model interpolation. Future applications may extend this to larger model families, different domains beyond vision, or more complex stitching patterns beyond binary configurations, though scaling challenges remain.

• →KLAS automates neural network stitching selection using KL divergence similarity metrics rather than heuristics, solving combinatorial complexity issues.
• →The framework achieves up to 1.21% higher ImageNet-1K accuracy at equivalent computational cost or enables 1.33x FLOPs reduction at maintained accuracy.
• →The approach generalizes across model families, unlike existing methods that require family-specific manual configuration.
• →Practical impact for edge computing and resource-constrained environments where flexible model selection optimizes deployment costs.
• →KL divergence-based similarity proves more effective than heuristic approaches for identifying optimal model stitching configurations.