Sharing Knowledge for Meta-learning with Feature Descriptions
Authors: Tomoharu Iwata, Atsutoshi Kumagai
NeurIPS 2022 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In our experiments, we demonstrate that the proposed method achieves better predictive performance than the existing meta-learning methods using a wide variety of real-world datasets provided by the statistical office of the EU and Japan. |
| Researcher Affiliation | Industry | Tomoharu Iwata NTT Communication Science Laboratories Kyoto, Japan EMAIL Atsutoshi Kumagai NTT Computer and Data Science Laboratories Tokyo, Japan EMAIL |
| Pseudocode | Yes | Algorithm 1 Meta-learning procedure of our model. |
| Open Source Code | No | The code is proprietary. |
| Open Datasets | Yes | For evaluating the proposed method, we used the following two data: e-Stat and Eurostat. ... The e-Stat data were obtained from the official statistics of Japan using API 1. The Eurostat data were obtained from the statistical office of the European Union using API 2. (Footnotes 1 and 2 provide URLs: 1https://www.e-stat.go.jp/en. 2https://ec.europa.eu/eurostat/data/database.) |
| Dataset Splits | Yes | For each of the e-Stat and Eurostat data, we sampled 700 datasets for meta-training, 100 for meta-validation, and 200 for meta-test. |
| Hardware Specification | No | Table 7 shows the average computation time in seconds for meta-training on computers with 2.60GHz CPUs. This is not a specific enough hardware detail (e.g., no CPU model or other components). |
| Software Dependencies | No | We implemented the proposed method with Py Torch [15]. This mentions PyTorch but does not specify a version number. |
| Experiment Setup | Yes | In the proposed model, we used a three-layered feed-forward neural network with 128 hidden and output units for the sentence, feature, and instance encoders, f SE, f FE, and f IE, and a three-layered feed-forward neural network with 128 hidden units and a single output unit for mean function g. For the activation function, we used rectified linear unit Re LU(x) = max(0, x). For GP, we used RBF kernels, k(z, z ) = α exp γ ||z - z ||^2 + βδ(z, z ), where α, β, and γ were kernel parameters to be meta-trained. We optimized our models using Adam [10] with learning rate 10^-3, batch dataset size 32, and dropout rate 0.1 [22]. The meta-validation datasets were used for early stopping, for which the maximum number of meta-training epochs was 5,000. |