Learning to Explore for Stochastic Gradient MCMC
Authors: Seunghyun Kim, Seohyeon Jung, Seonghyeon Kim, Juho Lee
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we will evaluate the performance of L2E in various aspects. Through extensive experiments, we would like to demonstrate followings: L2E shows excellent performance on real-world image classification tasks and seamlessly generalizes to the tasks not seen during meta-training. L2E produces similar predictive distribution to HMC as well as good mixing of BNNs posterior distribution in weight space. L2E can effectively explore and sample from BNNs posterior distribution, collecting diverse set of parameters both in weight space and function space. Experimental details In this paragraph, we explain our experimental settings. We evaluate L2E and baseline methods on 4 datasets: fashion-MNIST (Xiao et al., 2017), CIFAR-10, CIFAR-100 (Krizhevsky et al., 2009) and Tiny Image Net (Le & Yang, 2015). |
| Researcher Affiliation | Collaboration | 1Kim Jaechul Graduate School of AI, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, South Korea 2Nexon Korea, Seongnam, South Korea(This work was conducted while the author was in KAIST.) 3AITRICS, Seoul, South Korea. |
| Pseudocode | Yes | Algorithm 1 Meta training procedure and Algorithm 2 Inner Loop |
| Open Source Code | Yes | Code is available at https://github.com/ ciao-seohyeon/l2e. |
| Open Datasets | Yes | We evaluate L2E and baseline methods on 4 datasets: fashion-MNIST (Xiao et al., 2017), CIFAR-10, CIFAR-100 (Krizhevsky et al., 2009) and Tiny Image Net (Le & Yang, 2015). |
| Dataset Splits | No | The paper mentions 'validation data point' in the context of meta-objective calculation, but it does not provide specific details on the dataset splits (percentages, counts, or explicit splitting methodology) for training, validation, and testing in the main experiments, or reference standard splits with citation. |
| Hardware Specification | Yes | We use NVIDIA RTX-3090 GPU with 24GB VRAM and NVIDIA RTX A6000 with 48GB for all experiments. This research was supported with Cloud TPUs from Google s TPU Research Cloud (TRC). |
| Software Dependencies | No | The paper states 'We use JAX library to conduct our experiments.' and mentions 'tensorflow dataset' but does not provide specific version numbers for these software components or any other key libraries. |
| Experiment Setup | Yes | We report hyperparameters for each methods in Table 12, Table 13 and Table 14. We tune the hyperparameters of methods using BMA NLL as criterion with number of 10 samples. For all methods including L2E, we tune learning rate(step size), weight decay(prior variance) and momentum decay term. |