MEGA Hub

TESLA: Taylor Expansion of Sinusoidal Learnable Activations

Authors

Do you know Daehwa Ko?You can claim authorship or link another user.Do you know Jaehyeon Kim?You can claim authorship or link another user.Do you know Seunghyun Ham?You can claim authorship or link another user.Do you know Jay Hoon Jung?You can claim authorship or link another user.

Abstract

The parity problem--deciding whether the number of ones in a binary vector is odd or even--remains challenging for standard neural networks due to linear inseparability and the need for global interactions. We propose TESLA, an activation defined as a learnable combination of sine and cosine terms, enabling explicit control over polynomial degree and selective amplification of high-order components. Theoretically, we show that constraining TESLA's coefficients yields Lipschitz/Rademacher complexity bounds and shapes the training dynamics to emphasize higher-frequency structure. Empirically, on parity with input length n = 32, TESLA attains strong generalization with 100K training samples (approximately 0.002% of the 2^32 input space) and remains robust under heavy corruption, retaining high accuracy with up to 30% label noise. We also compare against periodic and frequency-based baselines (SIREN, SNAKE, and Fourier feature embeddings) on parity and Forrelation. Beyond synthetic structure, TESLA delivers comparable performance on ImageNet-100, indicating that activation-level degree control transfers to more general vision workloads. Code: https://github.com/KAU-QuantumAILab/TESLA

Community

00

Publication notes

Author note
15 pages, 7 figures. Accepted at AISTATS 2026
Journal
Proceedings of the 29th International Conference on Artificial Intelligence and Statistics (AISTATS 2026), PMLR 300, 2026