New Framework Unifies and Enhances Deep Neural Network Perturbation Techniques

Researchers have introduced a unified framework for perturbing hidden activations in deep neural networks, a concept previously under-analyzed. This framework reveals that existing methods like Dropout and adversarial feature perturbation are specific forms of activation perturbation. The proposed method, Learning to Perturb Activations (LPA), adaptively perturbs activations using class-specific perturbations learned through Projected Gradient Descent (PGD). Experiments show LPA consistently outperforms existing techniques and complements other perturbation methods. AI

IMPACT Introduces a novel framework and method that could improve the generalizability and robustness of deep learning models.

RANK_REASON This is a research paper detailing a new framework and method for deep learning. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• adversarial feature perturbation
• Learning to Perturb Activations (LPA)
• Manifold Mixup

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →