Core Concepts
신규 클래스를 지속적으로 발견하고 이전 지식을 잊지 않으면서 모델을 향상시키는 새로운 패러다임인 적응형 발견 및 병합(ADM)을 제안합니다.
Stats
"Extensive experiments on several datasets show that ADM significantly outperforms existing class-incremental Novel Class Discovery (class-iNCD) approaches."
"The CIFAR10 and CIFAR100 datasets contain 50,000 and 10,000 32 × 32 color images for training and testing."
"The Tiny-ImageNet contains 100,000 images of 200 classes downsized to 64×64 colored images."
Quotes
"Extensive experiments on class-iNCD demonstrate that our method can significantly outperform the existing methods without increasing the computational cost."
"The proposed AFF and AMM preserve important base features."