2020 Fiscal Year Annual Research Report
A scalable privacy-preserving information retrieval system based on federated optimization, on-device intelligence and semantic matching
| Project/Area Number |
19H04215
|
|---|
| Research Institution | University of Tsukuba |
|---|
Principal Investigator |
于 海涛 筑波大学, 図書館情報メディア系, 助教 (30751052)
|
|---|
| Co-Investigator(Kenkyū-buntansha) |
吉川 正俊 京都大学, 情報学研究科, 教授 (30182736)
康 シン 徳島大学, 大学院社会産業理工学研究部(理工学域), 助教 (80777350)
|
|---|
| Project Period (FY) |
2019-04-01 – 2024-03-31
|
| Keywords | policy gradient / learning to rank / probabilistic regression |
|---|
| Outline of Annual Research Achievements |
This year our first task is the evaluation experiments on semantic matching. We found that recent ranking methods based on either reinforcement learning or adversarial learning boil down to policy gradient based optimization. Our results show that policy gradient based ranking methods are, by a large margin, inferior to many conventional ranking methods. The failures are largely attributable to the gradient estimation based on sampled rankings which significantly diverge from ideal rankings. Our second task is the development of a new framework for ranking based on deep probabilistic regression model. The prediction is formulated as a distribution rather than a deterministic score, which enables us to better deal with the inherent uncertainty with respect to the prediction output.
|
| Current Status of Research Progress |
Current Status of Research Progress
2: Research has progressed on the whole more than it was originally planned.
Reason
The fact that our planned research is being conduced smoothly can be demonstrated as follow: First, we finished an in-depth evaluation experiment on semantic matching, which covers popular techniques, such as reinforcement learning and adversarial learning. This echoes our proposal well. Second, we propose to perform ranking based on deep probabilistic regression. Thanks to this, it becomes easy for us to incorporate uncertainty analysis. Good estimation with uncertainty provides a valuable extra bit of information to establish trustworthiness with users. Moreover, it provides us insights on developing novel on-device intent-detection model, since stochastic noise is added during parameter aggregation.
|
| Strategy for Future Research Activity |
For the future work, our first objective is to conduct an in-depth evaluation of the designed on-device intent-detection model. The robustness across different kinds of devices will be taken into account. In particular, a series of user studies for testing the efficiency and effectiveness are planned. Our second objective is to develop effective federated optimization framework for collaboratively learning the shared intent-detection model.
|
