Learning representations that generalize systematically

点击翻译按钮获取中文摘要

Deep learning algorithms are the driving force behind virtually all recent excitement around AI. Despite this, there are clear signs that the performance of these methods deteriorates outside of laboratory conditions. I set out a research program to develop learning representations that generalize learned associations beyond the training distribution, which we refer to as systematic generalization. To this end, we explore two broad categories of learning algorithms: (1) self-supervised learning methods and (2) iterated learning. We also propose to mathematically formalize the notion of systematic generalization and theoretically analyze promising avenues to achieve it.