Doorkeeper

Imperfect Information Learning Team Seminar (Talk by Ming-Kun Xie, Nanjing University of Aeronautics and Astronautics).

2023-06-01(木)09:00 - 10:00 JST
オンライン リンクは参加者だけに表示されます。
申し込む

申し込み受付は終了しました

今後イベント情報を受け取る

参加費無料

詳細

This is an online seminar. Registration is required.
【 Imperfect Information Learning Team】
【Date】2023/June/1(Thu) 9:00-10:00(JST)
*【Speaker】Ming-Kun Xie, Nanjing University of Aeronautics and Astronautics *

Title: Weakly-supervised multi-label learning

Abstract:
Single-label supervised learning assumes that each instance is associated with only one class label, while many real-world scenarios can be multi-labeled, where each instance consists of multiple semantics. Multi-Label Learning (MLL) is a practical and effective paradigm for handling examples with multiple labels. Compared to single-label multi-class classification, precisely labeling examples in the MLL scenarios becomes much more difficult due to the increasing size of the label space. This leads to the following problems: 1) the labeling of training examples becomes expensive; 2) the labeling information may be corrupted due to unavoidable reasons. The former makes it hard to use MLL models in many realistic applications under the labeling cost constraint, while the latter would significantly degrade the generalization performance of the trained models. To address these challenges, my research focuses on training effective and robust MLL models with low labeling cost in complex realistic scenarios. Towards this goal, my research has made progress in extending the labeling strategies in the multi-label scenarios, including: 1) assigning a candidate label set to each instance; 2) providing a label pair with pairwise relevance ordering; 3) annotating only one positive label for each instance, and designing learning algorithms to improve the effectiveness and robustness of MLL under various types of weak supervision.

コミュニティについて

RIKEN AIP Public

RIKEN AIP Public

Public events of RIKEN Center for Advanced Intelligence Project (AIP)

メンバーになる