Learning Deep Kernels for Non-Parametric Two-Sample Tests
Authors: Feng Liu, Wenkai Xu, Jie Lu, Guangquan Zhang, Arthur Gretton, Danica J. Sutherland
ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In experiments, we establish the superior performance of our deep kernels in hypothesis testing on benchmark and real-world data. |
| Researcher Affiliation | Collaboration | 1Australian Artificial Intelligence Institute, University of Technology Sydney, Sydney, NSW, Australia 2Gatsby Computational Neuroscience Unit, University College London, London, UK 3Toyota Technological Institute at Chicago, Chicago, IL, USA. |
| Pseudocode | Yes | Algorithm 1 Testing with a learned deep kernel |
| Open Source Code | Yes | The code of our deep-kernel-based two sample tests is available at github.com/fengliu90/DK-for-TST. |
| Open Datasets | Yes | The MNIST dataset contains 70 000 handwritten digit images (Le Cun et al., 1998). CIFAR-10.1 (Recht et al., 2019) is an attempt to collect a new test set for the very popular CIFAR-10 image classification dataset (Krizhevsky, 2009). |
| Dataset Splits | No | The paper states 'Thus we split the data, get ktr arg maxk ˆJλ(Str P , Str Q ; k), then compute the test statistic and permutation threshold on Ste P , Ste Q using ktr.' This defines training and testing data (Str and Ste) but does not mention a distinct validation set. |
| Hardware Specification | Yes | FL, JL and GZ gratefully acknowledge the support of the NVIDIA Corporation with the donation of two NVIDIA TITAN V GPUs for this work. |
| Software Dependencies | No | The paper mentions using 'Adam optimizer (Kingma & Ba, 2015)' and implicitly uses deep learning frameworks, but does not provide specific version numbers for any software dependencies. |
| Experiment Setup | Yes | For Blob, HDGM and Higgs, φω is a five-layer fullyconnected neural network, with softplus activations. the number of neurons in hidden and output layers of φω are set to 50 for Blob, 3d for HDGM and 20 for Higgs, where d is the dimension of samples. We use the Adam optimizer (Kingma & Ba, 2015). |