Paper accepted as Spotlight at AISTATS 2026
2026-01-23
AISTATS 2026 Spotlight Presentation
High-Performance Self-Supervised Learning by Joint Training of Flow Matching
Kosuke Ukita, Tsuyoshi Okita
Our paper "High-Performance Self-Supervised Learning by Joint Training of Flow Matching" was accepted as a Spotlight presentation (Top 3%) at AISTATS 2026.
This work proposes a new foundation model for sensor data by jointly training a Flow Matching generative model so that it can handle both generation and recognition tasks. By leveraging the structure of conditional generation processes in diffusion models and Flow Matching, we build a dynamic conditioning mechanism that greatly reduces computational cost while achieving high performance for both signal generation and human activity recognition on downstream tasks.
- Conference: The 29th International Conference on Artificial Intelligence and Statistics (AISTATS 2026)
- Presentation type: Spotlight (top 3%)
- Location: Tangier, Morocco — May 2026
AISTATS 2026 · Spotlight
2026-01-23
AISTATS 2026 Spotlight Presentation
High-Performance Self-Supervised Learning by Joint Training of Flow Matching
Kosuke Ukita, Tsuyoshi Okita
我々の論文が国際会議AISTATS2026にSpotlight presentation (Top 3%)として採択されました.
本研究は,生成モデルであるFlow Matchingの表現学習能力を活用し,生成タスクと認識タスクを共同で達成できるようモデルを学習することで,センサデータのための新しい基盤モデルを提案するものです.
拡散モデルやFlow Matchingなど生成モデルが持つ条件付き生成プロセスの構造を活用し,動的な条件付け機構を構築することで,計算コストを大幅に削減しながら,下流タスクにおける信号生成と人間行動認識を高い精度で両立します.
精緻な細かい特徴表現を必要とする生成タスクに反して,認識タスクではグローバルな信号全体の情報を必要とし両立が困難とされています. 人間行動認識(HAR)という認識タスクにおいてSOTAを達成し,生成モデルを基盤モデルとして昇格させました.
- 会議名: The 29th International Conference on Artificial Intelligence and Statistics (AISTATS 2026)
- 発表形式: Spotlight(top 3%)
- 開催地: Tangier, Morocco — May 2026
