A Hierarchical Bayesian Model for Few-Shot Meta Learning
Authors: Minyoung Kim, Timothy Hospedales
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirical results demonstrating improved accuracy and calibration performance on both classification and regression benchmarks. and We show empirically that our HBM provides improved performance and calibration in all of these cases |
| Researcher Affiliation | Collaboration | Minyoung Kim1 & Timothy M. Hospedales1,2 1Samsung AI Center Cambridge, UK 2University of Edinburgh, UK |
| Pseudocode | Yes | Algorithm 1 Our few-shot meta learning algorithm. and Our meta-test algorithm is also summarised in Alg. 2 (Appendix). |
| Open Source Code | Yes | Real codes for the synthetic Sine Line regression dataset and the large-scale Vi T are also attached in the Supplement Material to help understanding of our algorithm. |
| Open Datasets | Yes | For standard benchmark comparison using the popular Res Net backbones... we test our method on: mini Imagenet and tiered Image Net, The CIFAR-FS dataset is formed by splitting the original CIFAR-100..., Sine-Line dataset (Finn et al., 2018)., Object pose estimation on Shape Net datasets (Gao et al., 2022; Yin et al., 2020). |
| Dataset Splits | Yes | For the standard benchmarks with Conv Net/Res Net backbones... We follow the standard protocols of (Wang et al., 2019; Mangla et al., 2020; Zhang et al., 2021): With 64/16/20 and 391/97/160 train/validation/test class splits for mini Image Net and tiered Image Net datasets, respectively... and The CIFAR-FS dataset is formed by splitting the original CIFAR-100 into 64/16/20 train/validation/test classes. |
| Hardware Specification | No | The paper mentions 'GPU memory footprint' and 'Per-episode training time' and refers to running on 'real-world classification/regression datasets' but does not specify exact GPU models, CPU types, or other hardware specifications. |
| Software Dependencies | No | The paper mentions using 'Py Torch' and the 'Higher' library, but it does not specify the version numbers for these software dependencies, e.g., 'PyTorch 1.9'. |
| Experiment Setup | Yes | With the stochastic gradient descent (SGD) optimizer, we set momentum 0.9, weight decay 0.0001, and initial learning rate 0.01 for mini Image Net and 0.001 for tiered Image Net. We have learning rate schedule by reducing the learning rate by the factor of 0.1 at epoch 70. and For training, we run 100 epochs, each epoch comprised of 2000 episodes. and we have either 3 steps without burn-in (for large-scale backbones Vi T) or 5 steps with 2 burn-in steps (for smaller backbones Conv Net, Res Net-18, and CNP). |