Adversarial Metric Learning

Authors: Shuo Chen, Chen Gong, Jian Yang, Xiang Li, Yang Wei, Jun Li

IJCAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The experimental results on toy data and practical datasets clearly demonstrate the superiority of AML to representative state-of-the-art metric learning models. In this section, empirical investigations are conducted to validate the effectiveness of AML. In detail, we first visualize the mechanism of the proposed AML on a synthetic dataset. Then we compare the performance of the proposed method AML (Algorithm 1) with three classical metric learning methods (ITML [Davis et al., 2007], LMNN [Weinberger and Saul, 2009] and Flat Geo [Meyer et al., 2011]) and five state-of-the-art metric learning methods (RVML [Perrot and Habrard, 2015], GMML [Zadeh et al., 2016], ERML [Yang et al., 2016], DRML [Harandi et al., 2017], and DRIFT [Ye et al., 2017]) on seven benchmark classification datasets. Next, all methods are compared on three practical datasets related to face verification and image matching. Finally, the parametric sensitivity of AML is studied.
Researcher Affiliation Academia 1School of Computer Science & Engineering, Nanjing University of Science & Technology 2Key Laboratory of Intelligent Perception & Systems for High-Dimensional Information 3Jiangsu Key Laboratory of Image & Video Understanding for Social Security {shuochen, chen.gong, csjyang, xiang.li.implus, csywei}@njust.edu.cn, junl.mldl@gmail.com
Pseudocode Yes Algorithm 1 Solving AML in Eq. (6) via gradient descent.
Open Source Code No The paper does not provide any information about open-source code for the described methodology. There are no links to repositories or explicit statements about code availability.
Open Datasets Yes The datasets are from the well-known UCI repository [Asuncion and Newman, 2007], which include Breast-Cancer, Vehicle, German-Credit, Image-Segment, Isolet, Letters and MNIST. The Pub Fig face dataset [Nair and Hinton, 2010] consists of of 2 104 pairs of images belonging to 140 people, in which the first 80% pairs are selected for training and the rest are used for testing. Similar experiments are performed on the LFW face dataset [Huang et al., 2007] which includes 13233 unconstrained face images of 5749 individuals. The image matching dataset MVS [Brown et al., 2011] consists of 64 64 gray-scale image sampled from 3D reconstructions of the Statue of Liberty (LY), Notre Dame (ND) and Half Dome in Yosemite (YO).
Dataset Splits Yes In each trial, 80% of examples are randomly selected as the training examples, and the rest are used for testing. The Pub Fig face dataset [Nair and Hinton, 2010] consists of of 2 104 pairs of images belonging to 140 people, in which the first 80% pairs are selected for training and the rest are used for testing.
Hardware Specification No The paper does not provide any specific hardware details such as GPU models, CPU models, or memory specifications used for running the experiments.
Software Dependencies No The paper mentions features extracted by DSIFT and Siamese-CNN, but does not specify any software packages or libraries with version numbers (e.g., Python, PyTorch, TensorFlow, scikit-learn, etc.) used in the implementation or for experiments.
Experiment Setup Yes The step size ρ is recommended to be fixed to 0.001 in our experiments. The parameters in our method such as α and β are tuned by searching the grid {10 3, 10 2, , 103}. the training pairs are generated by randomly picking up 1000c(c 1) pairs among the training examples. we follow existing works [Xie and Xing, 2013; Zhan et al., 2016] and adopt the k-NN classifier (k = 5) based on the learned metrics to investigate the classification error rate.