Wisdom of the Ensemble: Improving Consistency of Deep Learning Models
Authors: Lijing Wang, Dipanjan Ghosh, Maria Gonzalez Diaz, Ahmed Farahat, Mahbubul Alam, Chetan Gupta, Jiangzhuo Chen, Madhav Marathe
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To validate the theory using three datasets and two state-of-the-art deep learning classifiers we also propose an efficient dynamic snapshot ensemble method and demonstrate its value. Code for our algorithm is available at https://github.com/christa60/dynens. |
| Researcher Affiliation | Collaboration | Lijing Wang University of Virignia lw8bn@virginia.edu Dipanjan Ghosh Hitachi America Ltd. dipanjan.ghosh@hal.hitachi.com Maria Teresa Gonzalez Diaz Hitachi America Ltd. teresa.gonzalezdiaz@hal.hitachi.com Ahmed Farahat Hitachi America Ltd. ahmed.farahat@hal.hitachi.com Mahbubul Alam Hitachi America Ltd. Mahbubul.Alam@hal.hitachi.com Chetan Gupta Hitachi America Ltd. chetan.gupta@hal.hitachi.com Jiangzhuo Chen University of Virignia chenj@virginia.edu Madhav Marathe University of Virignia marathe@virginia.edu |
| Pseudocode | Yes | Algorithm 1: Pseudocode of the dynamic snapshot ensemble (Dyn Snap) |
| Open Source Code | Yes | Code for our algorithm is available at https://github.com/christa60/dynens. |
| Open Datasets | Yes | We conduct experiments using three datasets and two state-of-the-art models. YAHOO!Answers [36] is a topic classification dataset with 10 output categories, 140K and 6K training and testing samples. CIFAR10 and CIFAR100 [23] are datasets with 10 and 100 output categories respectively, 50k and 10k color images as training and testing samples. |
| Dataset Splits | Yes | The dataset, models and hyper-parameters are shown in Table 1. Table 1: Data and Models ... Training ... Validation ... Testing |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, memory amounts) used for running its experiments. |
| Software Dependencies | No | The paper does not list specific software components with their version numbers required for reproducibility. |
| Experiment Setup | Yes | The experiment settings for Single Base models are shown in Table 1. We set m = 20 for ensemble methods, and N = 10, β = β for Dyn Snap-cyc and Dyn Snap-step, Fd(t) in Dyn Snap-step is 1e 1, 1e 2, 1e 3 at 80, 120, 160 epochs, dropout with 0.1 drop probability. |