Hierarchical Variational Memory for Few-shot Learning Across Domains
Authors: Yingjun Du, Xiantong Zhen, Ling Shao, Cees G. M. Snoek
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct thorough ablation studies to demonstrate the effectiveness of each component in our model. The new state-of-the-art performance on cross-domain and competitive performance on traditional few-shot classification further substantiates the benefit of hierarchical variational memory. |
| Researcher Affiliation | Collaboration | Yingjun Du1, Xiantong Zhen1,2, Ling Shao3, Cees G. M. Snoek1 1AIM Lab, University of Amsterdam 2Inception Institute of Artificial Intelligence 3National Center for Artificial Intelligence, Saudi Data and Artificial Intelligence Authority |
| Pseudocode | No | The paper does not contain any pseudocode or clearly labeled algorithm blocks. |
| Open Source Code | Yes | Our code will be publicly released. 1https://github.com/YDU-uva/Hier Memory. |
| Open Datasets | Yes | We apply our method to four cross-domain few-shot challenges and two within-domain few-shot image classification benchmarks. Sample images from all datasets are provided in the appendix A. Cross-domain datasets The 5-way 5-shot cross-domain few-shot classification experiments use mini Imagenet (Vinyals et al., 2016) as training domain and test on four different domains, i.e., Crop Disease (Mohanty et al., 2016) containing plant disease images, Euro SAT (Helber et al., 2019) consisting of a collection of satellite images, ISIC2018 (Tschandl et al., 2018) containing dermoscopic images of skin lesions, and Chest X (Wang et al., 2017), a set of X-ray images. Within-domain datasets The traditional few-shot within-domain experiments are conducted on mini Imagenet (Vinyals et al., 2016) which consists of 100 randomly chosen classes from ILSVRC2012 (Russakovsky et al., 2015), and tiered Imagenet (Ren et al., 2019) which is composed of 608 classes grouped in 34 high-level categories. |
| Dataset Splits | Yes | Tasks are drawn from a dataset by randomly sampling a subset of classes, sampling points from these classes, and then partitioning the points into support and query sets. |
| Hardware Specification | No | The paper does not provide specific hardware details (such as GPU or CPU models, memory, or specific computing infrastructure) used for running the experiments. |
| Software Dependencies | No | The paper does not provide specific software dependencies with version numbers. |
| Experiment Setup | No | The paper describes model architectures and general configurations (e.g., "Res Net-10 backbone", "two-layer inference network"), but does not provide specific hyperparameters like learning rates, batch sizes, optimizers, or number of epochs in the main text. |