Permutation Invariant Learning with High-Dimensional Particle Filters
Fuente:
arXiv
Gespeichert in:
| Hauptverfasser: | , , , , , |
|---|---|
| Format: | Preprint |
| Veröffentlicht: |
2024
|
| Schlagworte: | |
| Online-Zugang: | |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866909371833778176 |
|---|---|
| author | Boopathy, Akhilan Muppidi, Aneesh Yang, Peggy Iyer, Abhiram Yue, William Fiete, Ila |
| author_facet | Boopathy, Akhilan Muppidi, Aneesh Yang, Peggy Iyer, Abhiram Yue, William Fiete, Ila |
| contents | Sequential learning in deep models often suffers from challenges such as catastrophic forgetting and loss of plasticity, largely due to the permutation dependence of gradient-based algorithms, where the order of training data impacts the learning outcome. In this work, we introduce a novel permutation-invariant learning framework based on high-dimensional particle filters. We theoretically demonstrate that particle filters are invariant to the sequential ordering of training minibatches or tasks, offering a principled solution to mitigate catastrophic forgetting and loss-of-plasticity. We develop an efficient particle filter for optimizing high-dimensional models, combining the strengths of Bayesian methods with gradient-based optimization. Through extensive experiments on continual supervised and reinforcement learning benchmarks, including SplitMNIST, SplitCIFAR100, and ProcGen, we empirically show that our method consistently improves performance, while reducing variance compared to standard baselines. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2410_22695 |
| institution | arXiv |
| publishDate | 2024 |
| record_format | arxiv |
| spellingShingle | Permutation Invariant Learning with High-Dimensional Particle Filters Boopathy, Akhilan Muppidi, Aneesh Yang, Peggy Iyer, Abhiram Yue, William Fiete, Ila Machine Learning Artificial Intelligence Sequential learning in deep models often suffers from challenges such as catastrophic forgetting and loss of plasticity, largely due to the permutation dependence of gradient-based algorithms, where the order of training data impacts the learning outcome. In this work, we introduce a novel permutation-invariant learning framework based on high-dimensional particle filters. We theoretically demonstrate that particle filters are invariant to the sequential ordering of training minibatches or tasks, offering a principled solution to mitigate catastrophic forgetting and loss-of-plasticity. We develop an efficient particle filter for optimizing high-dimensional models, combining the strengths of Bayesian methods with gradient-based optimization. Through extensive experiments on continual supervised and reinforcement learning benchmarks, including SplitMNIST, SplitCIFAR100, and ProcGen, we empirically show that our method consistently improves performance, while reducing variance compared to standard baselines. |
| title | Permutation Invariant Learning with High-Dimensional Particle Filters |
| topic | Machine Learning Artificial Intelligence |
| url | https://arxiv.org/abs/2410.22695 |