Towards Diverse Perspective Learning with Switch over Multiple Temporal Pooling

Jihyeon Seong, JUNGMIN KIM, Jaesik Choi

Keywords: timeseries classification, temporal pooling, temporal relationship, perspective learning

Abstract: Pooling is a widely used method for classification problems. In particular, poolings that consider temporal relationships have been proposed in the time series classification (TSC) domain. However, we found that there exists a data dependency on temporal poolings. Since each pooling has only one perspective, existing temporal poolings cannot solve data dependency problem with a fixed perspective learning. In this paper, we propose a novel pooling architecture for diverse perspective learning: switch over multiple pooling (SoM-TP). The massive case study using layer-wise relevance propagation (LRP) reveals the distinct view that each pooling has and ultimately emphasizes the necessity of diverse perspective learning. Therefore, SoM-TP dynamically selects temporal poolings according to time series data characteristics. The ablation study on SoM-TP shows how diverse perspective learning is achieved. Furthermore, pooling classification is investigated through input attribution by LRP. Extensive experiments are done with the UCR/UEA repository.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Deep Learning and representational learning