核心概念
Generative models like GeNIe leverage diffusion techniques to create challenging samples for deep learning models through data augmentation.
要約
Directory:
Abstract
Introduction
Proposed Method: GeNIe
Algorithm 1: GeNIe-Ada
Experiments
Few-shot Classification
Long-Tailed Classification
Fine-grained Few-shot Classification
Ablation and Analysis
Related Work
Abstract:
Data augmentation is crucial for training deep models to prevent overfitting.
GeNIe introduces a novel augmentation method using diffusion models.
GeNIe generates hard negative samples by blending source images with text prompts.
Introduction:
Augmentation enhances model robustness and adaptability.
Traditional methods like MixUp and CutMix may not produce natural images.
Generative models like GeNIe use diffusion techniques for data augmentation.
Proposed Method: GeNIe:
GeNIe leverages latent diffusion models conditioned on text prompts.
It generates hard negative samples by blending source images with target categories.
Algorithm 1: GeNIe-Ada:
Adaptive noise level selection strategy improves performance in various scenarios.
Experiments:
Few-shot Classification:
Evaluates the impact of GeNIe on few-shot classification settings.
Long-Tailed Classification:
Assesses the effectiveness of GeNIe in long-tailed data settings.
Fine-grained Few-shot Classification:
Compares the performance of GeNIe with other text-based augmentation techniques.
Ablation and Analysis:
Analyzes the semantic shift from source to target class using different noise levels.
Discusses the label consistency of generated samples and the effect of noise on model performance.
Related Work:
Discusses data augmentation, generative models, language-guided recognition models, and few-shot learning in relation to GeNIe's approach.
統計
データ拡張は深層モデルの過学習を防ぐために重要です。
最近の進歩により、拡散モデルなどの生成AIがより洗練された拡張技術を可能にしました。