Self-Paced Co-training
Authors: Fan Ma, Deyu Meng, Qi Xie, Zina Li, Xuanyi Dong
ICML 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results substantiate the superiority of the proposed method as compared with current state-of-the-art co-training methods. (Abstract) |
| Researcher Affiliation | Academia | 1Xi an Jiaotong University, Xi an, China 2University of Technology Sydney, Sydney, Australia. |
| Pseudocode | Yes | Algorithm 1 Alternative Optimization Algorithm for Solving SPa Co Model |
| Open Source Code | No | The paper does not contain an explicit statement or link indicating that the source code for the described methodology is publicly available. |
| Open Datasets | Yes | Course data: This data set contains 1,051 home pages collected from web sites of Computer Science departments of Cornell University2. 2Data available at http://www.cs.cmu.edu/afs/cs/project/theo20/www/data/; Advertisement data: This data set contains advertising images in web pages3. 3Data available at https://archive.ics.uci.edu/ml/datasets/Internet +Advertisements; Newsgroup data: This data set is related to 16 newsgroups postings from the Mini-Newsgroup data4. 4Data available at http://www.cs.cmu.edu/afs/cs/project/theo11/www/ naive-bayes/mini newsgroup.tar.gz.; We employ the Market-1501 set in our experiment. This data set contains 32,668 detected person bounding boxes of 1,501 identities (Zheng et al., 2015). |
| Dataset Splits | Yes | For each data set, 25% of the data are retained as test examples while the rest are used as training examples, i.e., including both labeled and unlabeled examples. Among the training samples, we choose 2k positive and 3 2k negative, k = 1, 2, 3, labeled instances for Course, 2k positive and 6 2k negative labeled examples for Advertisement, and 2 2k+1 positive and 2 2k negative labeled examples for Newsgroup, based on their different data size. ... In the experiments, 20% instances of training data with their labels are chose with labels, the rest of data are treated as unlabeled instances. Instead of directly selecting labeled samples from the whole data, we randomly sample 20% labeled samples for each class. |
| Hardware Specification | No | The paper mentions using deep learning networks (Alexnet, Googlenet, Vggnet) and SVMs, but does not specify any hardware details such as GPU/CPU models or memory used for training or experimentation. |
| Software Dependencies | No | The paper mentions using "SVM" as a base classifier and refers to "any off-the-shelf SVM toolbox" and "deep learning network" but does not provide specific software names with version numbers for reproducibility. |
| Experiment Setup | Yes | For our SPa Co algorithm, instead of tuning λ directly, we increase the number of nonzero element of v(j) in each training round. Besides, to judge unlabeled instances based on two views, we easily set γ as 1 throughout all our experiments. Its setting actually is not sensitive to the final performance of our algorithm. ... Over all experiments, parameters of each model are set following the training setting as (Zheng et al., 2016). For the SPa Co algorithm, in every iteration, the number of selected unlabeled instances is ranged from 1000 to 2000. |