Learning with Unpaired Data

Summary form only given, as follows. The complete presentation was not made available for publication as part of the conference proceedings. Many learning tasks can be summarized as learning a mapping from a structured input to a structured output, such as machine translation, image captioning, and image dehazing. Such mappings are usually learned on paired training data, where an input sample and its corresponding output are both provided. Collecting paired training data often involves expensive human annotation, and the scale of paired training data is therefore often limited. As a result, the generalization ability of models trained on paired data is also limited. One way to mitigate this issue is learning with unpaired data, which is far less expensive to collect. Taking machine translation as an example, the unpaired training data can be collected separately from newspapers in the source language and target language without any annotation. The challenge of unpaired learning turns into how to align the unpaired data. With carefully designed objectives, unpaired learning has achieved remarkable progress on several tasks. This talk will cover the data collection and training methods of several unpaired learning tasks to illustrate the power of learning with unpaired data.