| 研究課題/領域番号 |
25540100
|
|---|
| 研究種目 |
挑戦的萌芽研究
|
| 研究機関 | 京都大学 |
|---|
研究代表者 |
CUTURI Marco 京都大学, 情報学研究科, 准教授 (80597344)
|
|---|
| 研究期間 (年度) |
2013-04-01 – 2015-03-31
|
| キーワード | metric learning / kernel methods / probability simplex / empirical Bayes |
|---|
| 研究概要 |
二つの新しい手法を発表しました。第一は、機械学習エジア会議（Asian Conference on Machine Learning 2013, Canberra, Australia) 2013年度に、Generalized Aitchison Embeddings for Histograms (T.Le, M.Cuturi)を発表しました.第二は、Neural Information Processing Systems (NIPS)会議2013年度に、Sinkhorn Distances: Lightspeed Computation of Optimal Transport(M. Cuturi)を発表しました.
This research project has produced so far a new principled way to learn metrics for histograms using geometrical arguments. These findings were presented at the ACML conference (2013) and are currently under review for a possible publication in Machine Learning journal. As a side contribution of this project, we were also able to make significant progress in an adjacent field: that of proposing fast and accurate approximations for the optimal transport metric through a presentation at the NIPS conference (2013) in a spotlight talk (<6% accepted papers out of 1500 submissions). We have started the project in this second year by studying unsupervised approaches for metric and kernel learning for histograms that can leverage both ideas from information geometry and empirical bayesian calculus to propose novel metrics and kernels on histograms.
|
| 現在までの達成度 (区分) |
現在までの達成度 (区分)
1: 当初の計画以上に進展している
理由
一流会議で２回発表できまして、そして一流の雑誌に提出出来まして、計画以上に進展しています。
Our project has so far produced 2 conference papers (including one in the best conference in our field) and 1 journal paper under submission. We believe the project is progressing very smoothly and expect additional outcomes later this year.
|
| 今後の研究の推進方策 |
来年度には、ベイズに用いる計算に頼って、今よりも計算をGPUに移る事。
Our plan for the next and final year of this project will be to add more Bayesian perspectives to our current research agenda and establish a bridge between our current work, based on metrics, towards our final goal of designing kernels. Our computational approach will increasingly rely on GPGPUs, which we have started to master from last year. We thus expect additional investments in GPGPU computing capacities this year.
|
| 次年度の研究費の使用計画 |
会議に発表出来る様に、又海外にいる協力者と研究を進ませる様に、旅費を今年度とも使う予定があります。すでGPUの２枚をこの科研費の陰で購入出来ましたが、研究を進ませるように後２枚を買う必要になりました。
GPUの計算をうまく使える様になりまして、今年もGPUの２枚を購入する予定があります。尚、今年も夏の時期にヨーロッパに滞在する予定があります。
|
