2022 Fiscal Year Annual Research 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, スタッフサイエンティスト (80869206)
|
|---|
| Project Period (FY) |
2020-04-01 – 2023-03-31
|
| Keywords | free-energy / sequence prediction / recurrent neural network / robot / working memory / multi-modal / planning |
|---|
| Outline of Annual Research Achievements |
The project aimed for implementation of recurrent neural networks for sequence prediction with explicit memory (one or multi-slot). Training and planning for novel sequences was performed under the free-energy principle.
Experiments of previous FYs showed content agnostic information processing for visual modalities. This means that the model developed(self-organized) an efficient information processing strategy, in which image pixels are not predicted directly, instead, the network learned to generate control signals to store/read data from memory, attention, and transformation modules. As a result, strong generalization from small training data sets could be achieved.
Current experiments extended this concept to proprioceptive sequences with additional language generation.The experiments confirmed a self-organized utilization of the provided memory under training with back-propagation in this case as well, and the generalization performance from small training data sets was significantly higher for models with explicit memory.
Further, tactile sensor devices were developed to support ongoing/future studies that investigate online replanning under consideration of multi-modal predictions.
Each gripper of a humanoid robot was equipped with 10 sensor pads (6x palm, 4x tips).
In preparation of future experiments, a further robotic platform was designed and implemented: a 12 DoF, torque controlled, quadruped robot. It will be used to evaluate planning of gait patterns under consideration of prediction error during minimization of free-energy of sequential models.
|
