| 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 | Synaptic Pruning / Quality-Diversity / Information theory / Network metrics / Evolutionary Strategy |
|---|
| 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 year, we used an exploratory approach based on Quality-Diversity (QD) algorithms to perform a systematic analysis of the effects of pruning in neural models during evolution. We relied on mathematical tools from network science to capture network structure regularities and information-theoretic analysis and describe the learning process.
We focused on reinforcement learning problems (bipedal walker, lunar lander) and compared the results to a purely performance-based optimization technique. We also evaluated setups where a pruning operator evolves along the models.
Results from the analysis showed the emergence of patterns and regimes in the mutual information and the estimation of network measures. This exploratory work will provide a solid ground for guiding and facilitating the development of pruning algorithms.
A short paper describing our results has been accepted at the GECCO international conference, and a second one was submitted to the ALIFE conference.
|
| Current Status of Research Progress |
Current Status of Research Progress
2: Research has progressed on the whole more than it was originally planned.
Reason
Research progressed according to plan. We implemented the exploration algorithm, applied it to standard benchmarks in the reinforcement learning field (bipedal walker, lunar lander), and analyzed the results. Across all metrics considered, we found some that correlated with an acceleration of the learning rate (Louvain, assortativity) and are great candidates for direct pruning algorithms, as we hoped. Results have been summarized in research papers and submitted to top international conferences.
|
| Strategy for Future Research Activity |
Future work will test the approach in supervised learning (LeCun et al., 2015), using computer vision tasks solved with widely used and tested DNN models trained with state-of-the-art learning methods (Khan et al., 2022). This would facilitate the assessment of our method as it enables a controlled comparison with other pruning methods. Avoiding reinforcement learning scenarios will likely facilitate the evaluation of the pruning algorithm during its development, as this class of problems is notoriously unstable and difficult to solve. We also plan to use a larger set of metrics, with the only constraint being that they apply to directed acyclic graphs without considering the weights of the edges. Another approach is to consider each layer pairing as a bipartite graph to expand the set of measures that are potential candidates for fitness functions in the pruning phase.
|
