Task Detection in Continual Learning via Familiarity Autoencoders
Task Detection in Continual Learning via Familiarity Autoencoders
复制标题
DOI：
10.1109/smc53654.2022.9945326
复制
发表时间：
2022-10
期刊：
影响因子：
--
通讯作者：
Maxwell J. Jacobson;Case Q. Wright;Nan Jiang;Gustavo Rodriguez-Rivera;Yexiang Xue
中科院分区：
文献类型：
--
作者：
Maxwell J. Jacobson;Case Q. Wright;Nan Jiang;Gustavo Rodriguez-Rivera;Yexiang Xue
持续学习需要能够可靠地将以前学到的知识转移到新的任务中，而不会破坏已有的能力。进步神经网络[1]等方法实现了高质量的迁移学习，同时消除了潜在的灾难性遗忘问题。然而，大多数基于模块的持续学习系统在操作过程中需要任务标签-这一限制限制了它们在许多任务指标不透明的现实世界条件下的应用。提出了一种任务检测器神经算法，在获取任务信息的同时保持对遗忘的免疫力。我们提出的任务检测器允许渐进神经网络(和许多类似的系统)在测试期间无需任务标签即可运行。我们的任务检测器是由熟悉度自动编码器构建的，它从输入数据中识别所需任务的性质。通过在视频游戏和自动图像修复中的实验，验证了该方法的通用性和有效性。我们的结果显示在所有领域的任务识别近乎完美(>；.99 F1)，在MinAtar中的奖励高于公布的单任务分数，并对受损的人脸图片进行真实的图像修复。我们的集成方法的性能与配备了地面真实任务标签的渐进式系统几乎相同
Continual learning requires the ability to reliably transfer previously learned knowledge to new tasks without disrupting established competencies. Methods such as Progressive Neural Network [1] accomplish high-quality transfer learning while nullifying the insidious problem of catastrophic forgetting. However, most module-based continual learning systems require task labels during operation – a constraint that limits their application in many real-world conditions where task indicators are opaque. This paper proposes a task detector neural algorithm to acquire task information while maintaining immunity to forgetting. Our proposed task detector allows progressive neural networks (and many similar systems) to operate without task labels during test time. Our task detector is built from familiarity autoencoders which recognize the nature of the required task from input data. We demonstrate the generality and effectiveness of our approach through experiments in video game playing and automated image repair. Our results show near-perfect task recognition in all domains (>.99 F1), rewards above published single-task scores in MinAtar, and realistic image repairs on damaged human face pictures. The performance of our integrated method is nearly identical to the progressive systems equipped with ground-truth task labels1