Non-Negative Bregman Divergence Minimization for Deep Direct Density Ratio Estimation
Authors: Masahiro Kato, Takeshi Teshima
ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through our experiments, the proposed methods show a favorable performance in inlier-based outlier detection. and 5. Experiments. We experimentally show how existing DRE methods fail and D3RE succeeds when using neural networks. |
| Researcher Affiliation | Collaboration | Masahiro Kato 1 Takeshi Teshima 2 1Cyber Agent Inc., Tokyo, Japan 2The University of Tokyo, Tokyo, Japan. |
| Pseudocode | Yes | Algorithm 1 D3RE |
| Open Source Code | Yes | A code of the conducted experiments is available at https: //github.com/MasaKat0/D3RE. |
| Open Datasets | Yes | We construct positive and negative datasets from the CIFAR-10 (Krizhevsky, 2009) dataset with 10 classes. and In addition to CIFAR-10, we use MNIST (Le Cun et al., 1998) and fashion-MNIST (FMNIST) (Xiao et al., 2017), both of which have 10 classes. |
| Dataset Splits | No | The paper describes training and test sets ('1,000 positive samples... to train the models' and '10,000 test samples'), but does not explicitly mention a separate validation split or how data is partitioned for validation. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU models, CPU types, or memory specifications used for running the experiments. It only mentions software used (PyTorch, Adam optimizer) and model architectures (CNN, LeNet, Wide ResNet). |
| Software Dependencies | No | The paper mentions software like 'Py Torch' and 'Adam optimizer', but it does not provide specific version numbers for these software dependencies, which is required for reproducibility. |
| Experiment Setup | Yes | The model is trained by the Adam optimizer (Kingma & Ba, 2015) without a weight decay and with the parameters (β1, β2, ϵ) fixed at the default values of the implementation in Py Torch (Paszke et al., 2019), namely (0.9, 0.999, 10−8). We report the results for two learning rates, 1 × 10−4 and 1 × 10−5. |