2020 Fiscal Year Research-status Report
Self-Organized Multi-Level Working Memories Facilitate Predictive Coding Based Action Panning
| Project/Area Number |
20K19901
|
|---|
| 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, 研究員 (80869206)
|
|---|
| Project Period (FY) |
2020-04-01 – 2022-03-31
|
| Keywords | Bio-inspired Learning / Working Memory / Self Organization / Robot Learning / Actuve Inference / Variable Binding |
|---|
| Outline of Annual Research Achievements |
The conducted work explores visual working memories (VWM) for AI systems.
By learning to manipulate content in a VWM instead of learning to represent the content directly, an improved generalization performance could be achieved:
The model is able to manipulate previously unseen objects (block stacking) in a scenario that requires goal-directed planning. Generalization for new colors of objects and applied textures have been explored.
The developed method was introduced under the term "content agnostic information processing" and has been accepted for publication in "Neural Computation" on Mar 18, 2021. Link: https://groups.oist.jp/sites/default/files/imce/u103429/ContentAgnosticInformationProcessingPreprint.pdf
Further, application in extended robotic scenarios and in VR have been prepared.
|
| Current Status of Research Progress |
Current Status of Research Progress
2: Research has progressed on the whole more than it was originally planned.
Reason
With respect to the initial Research Proposal, project progression had to face problems caused by the following reasons:
1.) Outbreak of novel coronavirus disease (COVID-19) delayed work in the laboratory, robot repair, meetings, and planned international travel.
2.) The reviewers of the paper required additionally experiments that evaluate the representations in the proposed model in case of generalization for new textures.
Currently, work is conducted in relation to WP2 (research proposal), as planned, extensions of the model to aim for further modalities have been sketched out, but are not yet implemented.
|
| Strategy for Future Research Activity |
The structure for continuation of the project is outlined in the following:
1. Implementation of the designed model extensions (generalization of concept of counting between multiple actions, research area "general AI"); 2. Evaluation on automatically generated test sets (simple toy-data); 3. Improvement of model design/concept; 4. Design/implementation of robotic experiment (real robot or(/and) VR); 5. Final evaluation (complex data set); 6. Documentation / Journal publication
As a result, it is expected that we can show the generalization of the concept of "content agnostic information processing" from the domain of vision to the domain of action primitives in combination with higher-level representations (language, e.g. 'touch the object N times).
|
| Causes of Carryover |
Because of the outbreak of the Novel Coronavirus (COVID-19), international travel could not be conducted as planned.
Further, due to the pandemic, the university was partially closed, this caused a rearrangement in the project plan. Closing and "state of emergency" affected robot operation and repair, therefore work was prepared in VR.
The robot experiments are scheduled for the coming fiscal year. The budget of FY2020 will be used in FY2021 to implement the experimental robotic scenario for "counting" experiments.
International travel was shifted from FY 2020 to FY2021 (planned in Winter 2021).
|
