Deep Convolutional Ranking for Multilabel Image Annotation

Authors: Yunchao Gong; Yangqing Jia; Thomas Leung; Alexander Toshev; Sergey Ioffe

ICLR 2014 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments on the NUS-WIDE dataset outperforms the conventional visual features by about 10%, obtaining the best reported performance in the literature.
Researcher Affiliation Collaboration Yunchao Gong UNC Chapel Hill yunchao@cs.unc.edu Yangqing Jia Google Research jiayq@google.com Thomas K. Leung Google Research leungt@google.com Alexander Toshev Google Research toshev@google.com Sergey Ioffe Google Research sioffe@google.com
Pseudocode No The paper includes mathematical equations for loss functions but does not provide any pseudocode or clearly labeled algorithm blocks.
Open Source Code No The paper does not provide any statement or link indicating that the source code for the described methodology is publicly available.
Open Datasets Yes We performed experiments on the largest publicly available multilabel dataset, NUS-WIDE [4].
Dataset Splits No The paper states: "We used a subset of 150,000 images for training and used the rest of the images for testing." It does not specify a separate validation set split.
Hardware Specification No The paper mentions training "on a cluster" but does not provide specific details about the hardware components (e.g., GPU/CPU models, memory).
Software Dependencies No The paper mentions software components and techniques (e.g., Dropout, k NN, SVM, RELU) but does not specify their version numbers.
Experiment Setup Yes The basic architecture of the network we use is similar to the one used in [20]. We use five convolutional layers and three densely connected layers. ... Convolution filter sizes are set to squares of size 11, 9, and 5... Each densely connected layer has output sizes of 4096. Dropout layers follow each of the densely connected layers with a dropout ratio of 0.6. ... The optimization of the whole network is achieved by asynchronized stochastic gradient descent with a momentum term with weight 0.9, with mini-batch size of 32. The global learning rate for the whole network is set to 0.002... For k NN, we set the bandwidth σ to 1 and k to 50... For SVM, we set the regularization parameter to C = 2...