Adaptive Sampling Probabilities for Non-Smooth Optimization
Authors: Hongseok Namkoong, Aman Sinha, Steve Yadlowsky, John C. Duchi
ICML 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We complement our analysis with experiments on several datasets. 4. Experiments We compare performance of our adaptive approach with stationary sampling distributions on real and synthetic datasets. |
| Researcher Affiliation | Academia | 1Management Science & Engineering, Stanford University, USA 2Electrical Engineering, Stanford University, USA 3Statistics, Stanford University, USA. Correspondence to: Hongseok Namkoong <hnamk@stanford.edu>, Aman Sinha <amans@stanford.edu>. |
| Pseudocode | Yes | Algorithm 1 Non-smooth Coordinate Descent; Algorithm 2 Stepsize Doubling Coordinate Descent; Algorithm 3 Coordinate Descent with Adaptive Sampling; Algorithm 4 KL Projection; Algorithm 5 Mirror Descent with Adaptive Sampling. |
| Open Source Code | No | The paper does not provide any explicit statements or links indicating that open-source code for the described methodology is available. |
| Open Datasets | Yes | We apply our method to two multi-class object detection datasets: Caltech-UCSD Birds-200-2011 (Wah et al., 2011) and ALOI (Geusebroek et al., 2005). |
| Dataset Splits | No | The paper mentions datasets used but does not explicitly specify the training, validation, and test splits with percentages, sample counts, or clear references to standard predefined splits used in their experiments. For example, for CUB-200-2011 and ALOI, it states 'n = 5994' and 'n = 108,000' respectively, but not how these were split. |
| Hardware Specification | No | The paper does not explicitly describe any specific hardware components (e.g., GPU/CPU models, memory details) used for running the experiments. |
| Software Dependencies | No | The paper does not list specific software dependencies with their version numbers (e.g., Python 3.x, PyTorch 1.x, or specific libraries with versions). |
| Experiment Setup | Yes | We run all experiments with pmin = 0.1/b and multiple values of c. We tune the stepsize x for both methods, using the form β/√t and tuning β. We set the blocksize as 50 features (b = 2110) and pmin = 0.01/b. We set pmin = 0.5/b to enforce enough exploration. We use softmax loss for CUB-200-2011 and a binary SVM loss for ALOI. We use X := {x ∈ Rm : kxk2 r} where r = 100 for CUB and r = 10 for ALOI. |