Deep Discriminative CNN with Temporal Ensembling for Ambiguously-Labeled Image Classification
Authors: Yao Yao, Jiehui Deng, Xiuhua Chen, Chen Gong, Jianxin Wu, Jian Yang12669-12676
AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we evaluate the performance of the proposed D2CNN approach on two synthesized datasets and two real-world datasets. |
| Researcher Affiliation | Academia | 1PCA Lab, the Key Laboratory of Intelligent Perception and Systems for High-Dimensional Information of Ministry of Education, School of Computer Science and Engineering, Nanjing University of Science and Technology, China 2Jiangsu Key Lab of Image and Video Understanding for Social Security 3National Key Laboratory for Novel Software Technology, Nanjing University, China |
| Pseudocode | No | No pseudocode or clearly labeled algorithm block was found. |
| Open Source Code | No | No statement regarding the release or availability of open-source code was found. |
| Open Datasets | Yes | Experiments on Fashion-Mnist dataset and SVHN dataset with manually added ambiguous labels. ... The Lost dataset is collected from the TV serial Lost by Cour et al. (Cour et al. 2009), which aims to associate the faces appear in some certain frames with the correct names captured from the corresponding subtitles. ... The Yahoo!News dataset contains the face images appeared in the news as well as the names (i.e. classes) in the corresponding captions. |
| Dataset Splits | Yes | The average classification accuracies of D2CNN and other baselines produced by the five-fold cross validation are shown in Table 1 |
| Hardware Specification | No | No specific hardware details (e.g., GPU/CPU models, memory, or cloud instances) used for experiments were mentioned. |
| Software Dependencies | No | The paper mentions 'Adam' as optimizer and 'Res Net' as backbone but does not provide specific version numbers for any software components or libraries. |
| Experiment Setup | Yes | For our D2CNN approach, we adopt a relatively shallow network Res Net-20 (He et al. 2016) as backbone... Adam (Kingma and Ba 2014) is utilized to optimize the networks with β1 = 0.9 and β2 = 0.999. Besides, we employ the weight decay of 0.0001, mini-batch size of 100, and ensembling weighting degree γ = 0.6 for all experiments. We train the networks for 300 epochs and enable T(t) to reach Tmax after 200 epochs in all runs. ... As for the proposed D2CNN approach, we employ the initial learning rate 0.001 and divide it by 1.25 after 100 and 200 epochs for all experiments on these synthesized datasets. The parameters α and Tmax are set to 0.001 and 100, respectively. |