2021 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, ポストドクトラルスカラー (80869206)
|
|---|
| Project Period (FY) |
2020-04-01 – 2023-03-31
|
| Keywords | working memory / content agnostic / Robot / free energy / goal directed planning |
|---|
| Outline of Annual Research Achievements |
1. Publication of work on visual working memory, "Emergence of Content-Agnostic Information Processing by a Robot Using Active Inference, Visual Attention, Working Memory, and Planning", in Journal "Neural Computation".
2. The concept of content-agnostic information processing has been implemented in the context of proprioception and language.
Results show that additional working memory results in hierarchical organization of internal representations of learned models.
As a further result, supporting evidence for an improved task performance of models utilizing a certain class of additional working memory for repetition/counting tasks could be found.
3. Design and implementation of a robotic platform (quadruped, 12DoF torque control) for evaluation of the developed methods.
|
| Current Status of Research Progress |
Current Status of Research Progress
3: Progress in research has been slightly delayed.
Reason
Restriction in relation to the situation caused by the novel Coronavirus affected project plan. Robot design and implementation was also impacted by chip-shortage caused by lock-downs.
Paper review process, longer than expected due to extension of conducted experiments.
|
| Strategy for Future Research Activity |
The future research will focus on:
First, extension of the current scenario to include known (remember from training) and unknown patterns (generate from demonstration; store in working memory). It is expected that the proposed model is able to represent two strategies (in terms of internal activation) that allows (1) generation of learned knowledge in great detail and (2) generation of novel samples with strong generalization.
Second, publication of the results of working memory for the counting scenario.
This includes findings of the internal organization of task the representation and performance comparisons between various model architectures.
Third, extension to real-world robotic scenario and under consideration of multitask problems as a potential line of future research.
|
| Causes of Carryover |
Delays in research plan because of pandemic situation of covid-19.
Design and implementation of robotic platform (quadruped robot, 12Dof).
|
