Shuai Li

Guoxia Wang, Shuai Li, Congliang Chen et al.

Large Language Models (LLMs) face increasing loss spikes during scaling, undermining training stability and final performance. While gradient clipping mitigates this issue, traditional global approaches poorly handle parameter-specific gradient variations and decaying gradient norms. We propose **AdaGC**, an adaptive gradient clipping framework that automatically adjusts local thresholds per parameter through exponential moving average of gradient norms. Theoretical analysis proves AdaGC's convergence under non-convex conditions. Extensive experiments demonstrate significant improvements: On Llama-2 7B/13B, AdaGC completely eliminates loss spikes while reducing WikiText perplexity by 3.5% (+0.14pp LAMBADA accuracy) for 7B and achieving 0.65% lower training loss with 1.47% reduced validation perplexity for 13B compared to global clipping. For CLIP ViT-Base, AdaGC converges 25% faster than StableAdamW with full spike elimination. The method shows universal effectiveness across architectures (Llama-2 7B/13B) and modalities (CLIP), with successful integration into diverse optimizers like AdamW and Lion. Source code will be released on GitHub.

Shuai Li, Xiaoguang Ma, Shancheng Jiang et al.

Remarkable successes were made in Medical Image Classification (MIC) recently, mainly due to wide applications of convolutional neural networks (CNNs). However, adversarial examples (AEs) exhibited imperceptible similarity with raw data, raising serious concerns on network robustness. Although adversarial training (AT), in responding to malevolent AEs, was recognized as an effective approach to improve robustness, it was challenging to overcome generalization decline of networks caused by the AT. In this paper, in order to reserve high generalization while improving robustness, we proposed a dynamic perturbation-adaptive adversarial training (DPAAT) method, which placed AT in a dynamic learning environment to generate adaptive data-level perturbations and provided a dynamically updated criterion by loss information collections to handle the disadvantage of fixed perturbation sizes in conventional AT methods and the dependence on external transference. Comprehensive testing on dermatology HAM10000 dataset showed that the DPAAT not only achieved better robustness improvement and generalization preservation but also significantly enhanced mean average precision and interpretability on various CNNs, indicating its great potential as a generic adversarial training method on the MIC.