کنفرانس مهندسی برق ایران

صفحه اصلی / سی و چهارمین کنفرانس بین المللی مهندسی برق

SPNP: A Calibration-Aware and Sparse Output Normalization for Deep Neural Networks

نویسندگان :

Morteza Taheri¹

1- دانشگاه زنجان

کلمات کلیدی :

SPNP،softmax replacement،probability calibration

چکیده :

In this paper, we introduce Shifted Power Normalized Probabilities (SPNP), a simple closed-form output activation function designed as an alternative to Softmax for probability normalization. The proposed method performs sample-wise standardization of logits, followed by a minimal shift and power-based normalization. SPNP improves probability calibration and enables controllable sparsity without modifying network architectures or the underlying loss formulation. Experiments on CIFAR-100, CIFAR-10, and MNIST, where all models are trained using the same standard cross-entropy loss applied to logits, demonstrate that SPNP significantly improves calibration metrics such as ECE, MCE, and the Brier score, while maintaining or improving classification accuracy. In addition, SPNP produces sparse output distributions and introduces negligible inference-time computational overhead, making it a practical and reliable alternative to Softmax for calibrated probabilistic inference.