Fast Context Adaptation via Meta-Learning
Authors: Luisa Zintgraf, Kyriacos Shiarli, Vitaly Kurin, Katja Hofmann, Shimon Whiteson
ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we empirically evaluate CAVIA on regression, classification, and RL tasks. |
| Researcher Affiliation | Collaboration | 1University of Oxford 2Latent Logic 3Microsoft Research. Correspondence to: Luisa Zintgraf <luisa.zintgraf@cs.ox.ac.uk>. |
| Pseudocode | No | The paper describes the methods using mathematical equations and natural language, but does not include structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code is available at https://github.com/lmzintgraf/cavia. |
| Open Datasets | Yes | We start with the regression problem of fitting sine curves from Finn et al. (2017a). ... We train CAVIA on the Celeb A (Liu et al., 2015) training set... Mini-Imagenet (Ravi & Larochelle, 2017). ... Mu Jo Co (Todorov et al., 2012) tasks using the setup of Finn et al. (2017a). |
| Dataset Splits | Yes | In few-shot learning problems, we are given distributions over training tasks ptrain(T ) and test tasks ptest(T ). ... For each such task, the model is updated using L or J and only a few data points (Dtrain or τ train). Performance of the updated model is reported on Dtest or τ test. ... The Mini-Imagenet dataset consists of 64 training classes, 12 validation classes, and 24 test classes. ... perform model selection on the validation set |
| Hardware Specification | Yes | The NVIDIA DGX-1 used for this research was donated by the NVIDIA corporation. |
| Software Dependencies | No | The paper mentions using common machine learning frameworks and algorithms (e.g., policy gradient, TRPO) but does not specify exact version numbers for any software dependencies like Python, PyTorch, TensorFlow, or specific libraries. |
| Experiment Setup | Yes | We use a neural network with two hidden layers and 40 nodes each. The number of context parameters varies between 2 and 50. Per meta-update we use a batch of 25 tasks. ... We chose an inner-learning rate of 1.0 without tuning. ... All our models were trained with two gradient steps in the inner loop and evaluated with two gradient steps. Following (Finn et al., 2017a), we ran each experiment for 60, 000 meta-iterations. |