Submodularity in Data Subset Selection and Active Learning
Authors: Kai Wei, Rishabh Iyer, Jeff Bilmes
ICML 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We extensively evaluate the proposed framework on text categorization and handwritten digit recognition tasks with four different classifiers, including deep neural network (DNN) based classifiers. Empirical results indicate that the proposed framework yields significant improvement over the state-of-the-art algorithms on all classifiers. |
| Researcher Affiliation | Academia | Kai Wei KAIWEI@U.WASHINGTON.EDU Rishabh Iyer RKIYER@U.WASHINGTON.EDU Jeff Bilmes BILMES@U.WASHINGTON.EDU University of Washington, Seattle, WA 98195, USA |
| Pseudocode | Yes | Algorithm 1 Filtered Active Submodular Selection |
| Open Source Code | No | The paper mentions using third-party tools like LIBLINEAR and Caffe, but does not provide any explicit statement or link for the source code of their own methodology. |
| Open Datasets | Yes | We evaluate text categorization on the 20 Newsgroups data set 1, which consists of 18774 articles divided almost evenly among 20 different Use Net discussion groups (Lang, 1995). ... We evaluate the handwritten digit recognition task on the MNIST database 2, which consists of 60,000 training and 10,000 test samples. |
| Dataset Splits | Yes | For each instance of the experiment, we randomly split 2/3 of the whole data set as the training and test samples. ... The MNIST database... consists of 60,000 training and 10,000 test samples. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running the experiments. |
| Software Dependencies | No | The paper mentions using 'LIBLINEAR tools' and 'Caffe' but does not specify their version numbers. |
| Experiment Setup | Yes | For mini-batch active learning experiments, we first randomly label B = 100 samples, on which we train a classifier as the initial model. In each iteration, additional B unlabeled examples are selected for labeling to update the model. We evaluate for T = 10 iterations ending with a total of k = 1000 labeled examples. ... For FASS, we fix βt = β = 4000, t and test four different submodular objectives: f NB, f NN, ffac, and ffs (c = 0.1). ... We apply a Laplace smoothing parameter of 0.02 for training all NB models in the experiments. ... A DNN model, which consists of two convolution layers followed by two fully connected layers, is trained using Caffe... |