| Project/Area Number |
20K19901
|
|---|
| Research Category |
Grant-in-Aid for Early-Career Scientists
|
| Allocation Type | Multi-year Fund |
|---|
| Review Section |
Basic Section 61050:Intelligent robotics-related
|
|---|
| Research Institution | Okinawa Institute of Science and Technology Graduate University |
|---|
Principal Investigator |
Queisser Jeffrey 沖縄科学技術大学院大学, 認知脳ロボティクス研究ユニット, スタッフサイエンティスト (80869206)
|
|---|
| Project Period (FY) |
2020-04-01 – 2023-03-31
|
| Project Status |
Completed (Fiscal Year 2022)
|
| Budget Amount *help |
¥2,080,000 (Direct Cost: ¥1,600,000、Indirect Cost: ¥480,000)
Fiscal Year 2021: ¥910,000 (Direct Cost: ¥700,000、Indirect Cost: ¥210,000)
Fiscal Year 2020: ¥1,170,000 (Direct Cost: ¥900,000、Indirect Cost: ¥270,000)
|
|---|
| Keywords | Generalization / Free Energy Minimization / Predictive Model / Robot / Learning / Planning / free-energy / sequence prediction / recurrent neural network / robot / working memory / multi-modal / planning / content agnostic / free energy / goal directed planning / Bio-inspired Learning / Working Memory / Self Organization / Robot Learning / Actuve Inference / Variable Binding |
|---|
| Outline of Research at the Start |
This project explores working memories (WM) for AI systems.
By learning to manipulate a WM instead of learning to represent the content directly, it is expected that a system is able to apply its knowledge to new situations without relearning and can find abstract representations of the world.
|
| Outline of Final Research Achievements |
The conducted work explored working memories (WM) for AI systems.
By learning to manipulate content in a visual WM instead of learning to represent the content directly, an improved generalization performance to unlearned situations could be achieved: The developed model is able to control a robot to manipulate previously unseen objects in a block stacking scenario that requires goal‐directed planning. Generalization was tested for new colors of objects and applied textures. Further, the concept of content-agnostic information processing was extended to the context of proprioception and language generation.
Results show that the introduced working memory modules result in hierarchical organization of internal representations of learned models without imposed explicit supporting constraints. As a further result, supporting evidence for an improved task performance of models utilizing a certain class of memory connectivity for repetition/counting tasks could be found.
|
| Academic Significance and Societal Importance of the Research Achievements |
This research increased sample efficiency and lowering the computational complexity for online robotic behavior generation, where “big-data” is not available and systems need efficient generalization from few observations.
Further, the models can be valuable for understanding brain (dis-)functions.
|
