Multimodal Model-Agnostic Meta-Learning via Task-Aware Modulation
Authors: Risto Vuorio, Shao-Hua Sun, Hexiang Hu, Joseph J. Lim
NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate the proposed model on a diverse set of few-shot learning tasks, including regression, image classification, and reinforcement learning. The results not only demonstrate the effectiveness of our model in modulating the meta-learned prior in response to the characteristics of tasks but also show that training on a multimodal distribution can produce an improvement over unimodal training.To investigate whether our method can acquire meta-learned prior parameters by learning tasks sampled from multimodal task distributions, we design and conduct experiments on a variety of domains, including regression, image classification, and reinforcement learning. The results demonstrate the effectiveness of our approach against other systems.The quantitative results are shown in Table 1. |
| Researcher Affiliation | Academia | Risto Vuorio 1 Shao-Hua Sun 2 Hexiang Hu2 Joseph J. Lim2 1University of Michigan 2University of Southern California vuoristo@gmail.com {shaohuas, hexiangh, limjj}@usc.edu |
| Pseudocode | Yes | Algorithm 1 MMAML META-TRAINING PROCEDURE. |
| Open Source Code | Yes | The code for this project is publicly available at https://vuoristo.github.io/MMAML. |
| Open Datasets | Yes | We evaluate our method (MMAML) and baselines in a variety of domains including regression, image classification, and reinforcement learning, under the multimodal task distributions.To create a multimodal few-shot image classification task, we combine multiple widely used datasets (OMNIGLOT [17], MINI-IMAGENET [34], FC100 [29], CUB [50], and AIRCRAFT [24]) to form a meta-dataset following the train/test splits used in the prior work, similar to [46]. |
| Dataset Splits | Yes | The dataset is split into meta-training and meta-testing sets, which are further divided into task-specific training Dtrain T and validation Dval T sets.A meta-learner learns about the underlying structure of the task distribution through training on the meta-training set and is evaluated on meta-testing set. The initialization of the parameters is trained by sampling mini-batches of tasks from D, computing the adapted parameters for all Dtrain T in the batch, evaluating adapted parameters to compute the validation losses on the Dval T and finally update the initial parameters θ using the gradients from the validation losses. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for running the experiments, such as GPU or CPU models. |
| Software Dependencies | No | The paper mentions the use of 'Mu Jo Co physics simulator [45]' but does not specify its version number or any other software dependencies with version information. |
| Experiment Setup | No | The paper explicitly states that 'Detailed network architectures and training hyper-parameters are different by the domain of applications, we defer the complete details to the supplementary material.' and 'Please refer to the supplementary materials for details and parameters for regression experiments.' This indicates that the specific experimental setup details are not present in the main text of the paper. |