Collaborative Research: AIS: Incremental Learning from Unbalanced Data in Nonstationary Environments

“该奖项是根据2009年美国复苏和再投资法案（公法111-5）资助的。”长期以来，计算智能的最终目标一直是通过从数据中发现和学习模式来模拟类脑智能。然而，在相关研究中，数据被认为是由一个潜在的固定的物理过程产生的。最近，新的算法出现了，可以容纳新的数据，或数据的不平衡分布。然而，从非平稳环境中学习，其中生成数据的底层过程随着时间的推移而变化，受到的关注较少，而在非平稳环境中学习的问题，增量提供不平衡数据几乎没有受到任何关注。既然大脑可以并且经常在这样的环境中学习，那么还需要一个通用的学习框架吗？适应呢？引入不平衡数据的非平稳环境怎么说都不为过。垃圾邮件检测、流行病学研究或气候变化分析只是此类场景的几个例子。考虑到这样一个数据不平衡的场景，这个项目的目标是开发一个通用框架，该框架可以识别是否以及何时发生了变化，学习新的内容，加强仍然相关的现有知识，并忘记可能不再相关的知识。我们的假设是，从不平衡和非平稳数据中学习可以通过战略性地使用i.)通过局部外推进行数据再生来实现。帮助平衡不平衡的数据集？结合ii.)一个增量生成的集成专家模型，使用动态分配的权重来模拟大脑的短期和长期记忆特性？以帮助跟踪变化的环境。

"This award is funded under the American Recovery and Reinvestment Act of 2009 (Public Law 111-5)"The ultimate goal of computational intelligence has long been emulating brain-like-intelligence by discovering and learning patterns from data. However, in related research, the data have been assumed to be generated by an underlying fixed physical process. Recently, new algorithms have emerged that can accommodate new data, or data with unbalanced distributions. However, learning from a non-stationary environment, where the underlying process that generates the data changes over time, has received less attention, whereas the problem of learning in a non-stationary environment that incrementally provides unbalanced data has received hardly any attention. Since the brain can and routinely does learn in such settings, the need for a general framework for learning from ? and adapting to ? a nonstationary environment that introduces unbalanced data can be hardly overstated. Spam detection, epidemiological studies, or analysis of climate change, are just a few examples of such scenarios. Given such a scenario of unbalanced data, the goal of this project is to develop a general framework that would recognize if and when there has been a change, learn novel content, reinforce existing knowledge that is still relevant, and forget what may no longer be relevant. Our hypothesis is that learning from unbalanced and nonstationary data can be achieved by strategic use of i.) data regeneration through local extrapolation ? to help balance the unbalanced dataset ? combined with ii.) an incrementally generated ensemble of experts model that use dynamically assigned weights to emulate short and long term memory properties of the brain ? to help track the changing environments.