Deep learning with evolutionary synaptic pruning
| Project/Area Number |
22KF0142
|
|---|
| Project/Area Number (Other) |
21F21768 (2021-2022)
|
|---|
| Research Category |
Grant-in-Aid for JSPS Fellows
|
| Allocation Type | Multi-year Fund (2023)
Single-year Grants (2021-2022) |
|---|
| Section | 外国 |
|---|
| Review Section |
Basic Section 61030:Intelligent informatics-related
|
|---|
| Research Institution | Ochanomizu University |
|---|
Principal Investigator |
オベル加藤 ナタナエル お茶の水女子大学, 基幹研究院, 講師 (10749659)
|
|---|
| Co-Investigator(Kenkyū-buntansha) |
DA ROLD FEDERICO お茶の水女子大学, 基幹研究院, 外国人特別研究員
|
|---|
| Project Period (FY) |
2023-03-08 – 2025-03-31
|
| Project Status |
Granted (Fiscal Year 2023)
|
| Budget Amount *help |
¥2,200,000 (Direct Cost: ¥2,200,000)
Fiscal Year 2024: ¥400,000 (Direct Cost: ¥400,000)
Fiscal Year 2023: ¥449,944 (Direct Cost: ¥449,944)
Fiscal Year 2022: ¥1,100,000 (Direct Cost: ¥1,100,000)
Fiscal Year 2021: ¥700,000 (Direct Cost: ¥700,000)
|
|---|
| Keywords | Evolutionary Strategy / Synaptic Pruning / Quality-Diversity |
|---|
| Outline of Research at the Start |
This project aims to develop a Quality-Diversity approach that dynamically prunes deep neural networks during exploration. The expected outcome of the project is an algorithmic framework for decreasing the number of parameters in deep learning models. We thus expect a decrease in energy consumption and computational requirements, with applications to embedded systems with low resources.
|
| Outline of Annual Research Achievements |
During this fiscal year, we developed a specialized Evolutionary Strategy optimization algorithm that prunes deep neural network models during the optimization process. The proposed approach combines quality diversity algorithms and pruning operators that achieve a model compression equivalent to state-of-the-art gradient-based pruning techniques. This work thus suggests the validity of gradient-free optimization methods as an alternative to machine learning pruning approaches to reinforcement learning problems. Results were submitted and published at The Genetic and Evolutionary Computation Conference (GECCO 2022), a top-tier international conference.
To further understand the interaction between pruning and parameters optimization, we devised an experiment implementing a population-based genetic algorithm. The initial population is initialized with randomly pruned neural models that successfully learn the task in the experiments described above.
We are attempting to find correlations between information-theoretic measures and sub-network that rapidly converge to a solution to the given task.
|
| 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 project implementation complies with the research plan with no significant delays for Work Package 1, which defined the workflow of the first year of the fellowship.
Task 1.1, "model selection and ES-GA training" was completed within month eight as expected.
The first objective of Task 1.2 is to refine the algorithm developed during the first months of the project implementation.
The technique utilized in the article published in GECCO 2022 is further refined by testing different genetic algorithm operators, such as selection criteria, mutation, and quality discriminators. Furthermore, the algorithm is applied to a larger set of reinforcement learning scenarios.
|
| Strategy for Future Research Activity |
The implementation of Work Package 1 will continue with the application of existing and novel benchmarks to test the energy efficiency of our method. The implementation of Work Package 2 will focus on a robotic scenario to further assess the performance of our approach to mode compression via synaptic pruning.
The first experiment will be based on a robotic arm performing simple tasks such as reaching and avoiding objects.
This scenario aims at porting the model used in the experiments utilized in the first Work Package to an agent with a more complex kinematic. To facilitate the learning process, vision will not be implemented, and the robot will perceive the environment with Cartesian coordinates.
The second robotic experiment will use a generative model to extrapolate information from a camera. Furthermore, the experimental setup will include challenging tasks such s as grasping and placing objects in a target area.
The expected outcome is a publication in a top-tier international conference and a submission to a highly reputed robotic journal.
|
Report
(2 results)
Research Products
(2 results)
