BOIL: Towards Representation Change for Few-shot Learning
Authors: Jaehoon Oh, Hyungjun Yoo, ChangHwan Kim, Se-Young Yun
ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | BOIL empirically shows significant performance improvement over MAML, particularly on cross-domain tasks. ... We validated both MAML/ANIL and BOIL on two general data sets, mini Image Net (Vinyals et al., 2016) and tiered Image Net (Ren et al., 2018), and two specific data sets, Cars (Krause et al., 2013) and CUB (Welinder et al., 2010). |
| Researcher Affiliation | Academia | 1Graduate School of Knowledge Service Engineering, KAIST 2Graduate School of Artificial Intelligence, KAIST {jaehoon.oh,yoohjun,kimbob,yunseyoung}@kaist.ac.kr |
| Pseudocode | No | No explicit 'Algorithm' block or pseudocode figure is present. The algorithm is described narratively with mathematical equations. |
| Open Source Code | No | All implementations are based on Torchmeta (Deleu et al., 2019) except for Warp Grad, and all results were reproduced according to our details. |
| Open Datasets | Yes | We validated both MAML/ANIL and BOIL on two general data sets, mini Image Net (Vinyals et al., 2016) and tiered Image Net (Ren et al., 2018), and two specific data sets, Cars (Krause et al., 2013) and CUB (Welinder et al., 2010). |
| Dataset Splits | Yes | We trained 4conv network and Res Net-12 for 30,000 and 10,000 epochs, respectively, and then used the model with the best accuracy on meta-validation data set to verify the performance. ... All the reported results are based on the model with the best validation accuracy. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU or CPU models used for running the experiments. |
| Software Dependencies | No | All implementations are based on Torchmeta (Deleu et al., 2019) except for Warp Grad. |
| Experiment Setup | Yes | We trained 4conv network and Res Net-12 for 30,000 and 10,000 epochs, respectively, and then used the model with the best accuracy on meta-validation data set to verify the performance. We applied an inner update once for both meta-training and meta-testing. The outer learning rate was set to 0.001 and 0.0006 and the inner learning rate was set to 0.5 and 0.3 for 4conv network and Res Net-12, respectively. |