Normalized training loss vs. baselines (Facebook).

<p>DCOR reports both reconstruction-only and total (with RLC); baselines report reconstruction-only. Each curve is normalized as in <a href="http://www.plosone.org/article/info:doi/10.1371/journal.pone.0335135#pone.0335135.e243" target="_blank">Eq (29)</a> by di...

Полное описание

Сохранить в:
Библиографические подробности
Главный автор: Hossein Rafieizadeh (22676722) (author)
Другие авторы: Hadi Zare (20073000) (author), Mohsen Ghassemi Parsa (22676725) (author), Hocine Cherifi (8177628) (author)
Опубликовано: 2025
Предметы:
Метки: Добавить метку
Нет меток, Требуется 1-ая метка записи!
Описание
Итог:<p>DCOR reports both reconstruction-only and total (with RLC); baselines report reconstruction-only. Each curve is normalized as in <a href="http://www.plosone.org/article/info:doi/10.1371/journal.pone.0335135#pone.0335135.e243" target="_blank">Eq (29)</a> by dividing by its epoch-1 value and EMA-smoothed (exponential moving average) with , where the EMA is computed as with . This normalization enables fair visual comparison across methods with different objectives and scales; the plot therefore emphasizes relative convergence trends (shape and stability) rather than raw magnitudes. Consistent with DCOR’s design, RLC regularizes late-phase training: the reconstruction curve decreases more conservatively than methods that minimize reconstruction alone, while the total objective continues to decrease.</p>