Learning Deep Latent Space for Multi-Label Classification

Authors: Chih-Kuan Yeh, Wei-Chieh Wu, Wei-Jen Ko, Yu-Chiang Frank Wang

AAAI 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments on multiple datasets with different scales confirm the effectiveness and robustness of our proposed method, which is shown to perform favorably against state-of-the-art methods for multi-label classification.
Researcher Affiliation Academia 1Research Center for IT Innovation, Academia Sinica, Taipei, Taiwan 2Department of Electrical Engineering, National Taiwan University, Taipei, Taiwan
Pseudocode Yes Algorithm 1: Learning of C2AE
Open Source Code No The paper does not contain any explicit statement or link indicating that the source code for the described methodology is openly available.
Open Datasets Yes To evaluate the performance of our proposed method, we consider the following datasets for experiments: iaprtc12, ESPGame, mirflickr, tmc2007, and NUS-WIDE. The first three datasets are image datasets used in (Guillaumin et al. 2009)
Dataset Splits Yes For NUS-WIDE, we follow the setting of (Gong et al. 2013) by discarding the instances with no positive labels and randomly select 150,000 instances for training and the remaining for testing. For fair comparisons with other CNN-based methods, we extract 4096-dimensional fc-7 feature for NUS-WIDE using a pre-trained Alex Net model. For the architecture of our C2AE, we have Fx composed of 2 layers of fully connected layers, while Fd and Fe are both single fully connected layer structures. For each fully connected layer, a total of 512 neurons are deployed. A leaky Re LU activation function is considered, while the batch size is fixed as 500. To select the parameters for C2AE, we randomly hold 1/6 of our training data for validation (with α selected from [0.1, 10] and λ fixed as 0.5).
Hardware Specification Yes To make additional remarks on the computation time, our C2AE only takes 10-15 minutes to perform training on NUSWIDE using a titan X GPU, which is much more efficient than training other DNN-based approaches, especially those require the learning of RNN.
Software Dependencies No The paper mentions using a 'pre-trained Alex Net model' but does not specify software dependencies with version numbers (e.g., Python, TensorFlow, PyTorch versions).
Experiment Setup Yes For the architecture of our C2AE, we have Fx composed of 2 layers of fully connected layers, while Fd and Fe are both single fully connected layer structures. For each fully connected layer, a total of 512 neurons are deployed. A leaky Re LU activation function is considered, while the batch size is fixed as 500. To select the parameters for C2AE, we randomly hold 1/6 of our training data for validation (with α selected from [0.1, 10] and λ fixed as 0.5).