Deep Metric Learning: The Generalization Analysis and an Adaptive Algorithm

Authors: Mengdi Huai, Hongfei Xue, Chenglin Miao, Liuyi Yao, Lu Su, Changyou Chen, Aidong Zhang

IJCAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We also conduct experiments on real-world datasets to verify the findings derived from the generalization error bound and demonstrate the effectiveness of the proposed adaptive DML method.
Researcher Affiliation Academia 1Department of Computer Science, University of Virginia, VA, USA 2Department of Computer Science and Engineering, SUNY at Buffalo, NY, USA
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide any concrete access to source code for the methodology described.
Open Datasets Yes Datasets. We adopt the following real-world datasets for our experiments: the MNIST 8v9 dataset1, the bone disease dataset2, and the wine quality dataset 3. 1http://yann.lecun.com/exdb/mnist/ 2https://sofonline.epi-ucsf.org/interface/ 3https://archive.ics.uci.edu/ml/datasets.php
Dataset Splits Yes For each dataset, 4,950 sample pairs are selected as the test set (no overlap with training set). Unless otherwise specified, the training data sizes for the MNIST 8v9 dataset, the bone disease dataset and the wine quality dataset are set as 160, 220 and 100, respectively. Correspondingly, the numbers of the generated training sample pairs for the MNIST 8v9 dataset, the bone dataset and the wine quality dataset are 12,720, 24,090 and 4,950, respectively. We do not use the validation set to tune parameters, but assign values by standard settings.
Hardware Specification Yes We implement the DML model using Google TensorFlow, and the training process is done locally using NVIDIA Ge Force GTX 1060 GPU.
Software Dependencies No The paper mentions 'Google TensorFlow' as the implementation tool but does not provide specific version numbers for TensorFlow or any other software libraries.
Experiment Setup Yes Unless otherwise specified, all the neural networks adopted in the experiments have 3 layers. For each dataset, the number of the units in each layer of the neural network is provided in Table 2... Additionally, Adam optimizer is used in the training process for DML and the learning rate is set as 1e-4. As for the activation function, we use ReLU because it is a 1-Lipschitz activation function and satisfies the Lipschitz-continuous condition.