Accelerating Natural Gradient with Higher-Order Invariance
Authors: Yang Song, Jiaming Song, Stefano Ermon
ICML 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 7. Experimental Evaluations In this section, we demonstrate the benefit of respecting higher-order invariance through experiments on synthetic optimization problems, deep neural net optimization tasks and policy optimization in deep reinforcement learning. |
| Researcher Affiliation | Academia | 1Computer Science Department, Stanford University. Correspondence to: Yang Song <yangsong@cs.stanford.edu>, Jiaming Song <tsong@cs.stanford.edu>, Stefano Ermon <ermon@cs.stanford.edu>. |
| Pseudocode | Yes | For geodesic correction, we only need to compute connection-vector products Γµ β γ γβ. This can be done with a similar idea to Hessian-vector products (Pearlmutter, 1994), for which we provide detailed derivations and pseudocodes in Appendix C. |
| Open Source Code | No | The paper references third-party code (OpenAI Baselines) but does not state that the authors' own source code for the methodology described in the paper is openly available. |
| Open Datasets | Yes | The datasets are CURVES, MNIST and FACES, all of which contain small gray-scale images of various objects, i.e., synthetic curves, hand-written digits and human faces. |
| Dataset Splits | No | The paper does not explicitly provide specific dataset split percentages, sample counts, or detailed splitting methodology for training, validation, and test sets in the main text. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., exact GPU/CPU models, processor types, or memory amounts) used for running its experiments. |
| Software Dependencies | No | The paper mentions software like 'TensorFlow' and 'OpenAI Gym' but does not provide specific version numbers for these or other key software components required for replication. |
| Experiment Setup | Yes | During training, and β are initialized to 1 and the learning rate is fixed to 0.5. |