引用iverjo · 2016-12
“How does this relate to Progressive Neural Networks [0]? That technique is also about accumulating knowledge (while not forgetting existing knowledge)”
拿它与新工作对比,同样是累积知识而不遗忘
用横向连接复用旧任务特征、避免灾难性遗忘的网络结构
提出渐进式网络:通过到先前已学特征的横向连接来利用旧知识,从而不会遗忘旧任务。在 Atari 和 3D 迷宫等强化学习任务上优于预训练加微调等常见基线,并用一种新的敏感度度量表明迁移同时发生在低层感知和高层控制层。
评论者在讨论“积累知识而不遗忘已有知识”的新方法时拿它做对照参考。
“How does this relate to Progressive Neural Networks [0]? That technique is also about accumulating knowledge (while not forgetting existing knowledge)”