How to Distribute Data across Tasks for Meta-Learning?
Authors: Alexandru Cioba, Michael Bromberg, Qian Wang, Ritwik Niyogi, Georgios Batzolis, Jezabel Garcia, Da-shan Shiu, Alberto Bernacchia6394-6401
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We prove these results mathematically on mixed linear regression, and we show empirically that the same results hold for few-shot image classification on CIFAR-FS and mini-Image Net. Our results provide guidance for allocating labels across tasks when collecting data for meta-learning. |
| Researcher Affiliation | Collaboration | Alexandru Cioba, 1 Michael Bromberg, 1 Qian Wang, 1 Ritwik Niyogi, 1 Georgios Batzolis, 2 Jezabel Garcia, 1 Da-shan Shiu, 1 Alberto Bernacchia 1 1 Media Tek Research, 2University of Cambridge |
| Pseudocode | No | The paper describes the algorithms in prose and mathematical equations but does not include structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | We prove these results mathematically on mixed linear regression, and we show empirically that the same results hold for few-shot image classification on CIFAR-FS and mini-Image Net (code in the supplementary material). |
| Open Datasets | Yes | We investigate the CIFAR-FS (Bertinetto et al. 2019) and mini-Image Net (Vinyals et al. 2017) datasets, which are fewshot versions of CIFAR-100 and Image Net, respectively. |
| Dataset Splits | Yes | In the meta-training phase, m tasks (τi)m i=1 are sampled from p(τ) and, for each task, nt i training data points Dt i = (xt ij, yt ij)nt i j=1 and nv i validation data points Dv i = (xv ij, yv ij)nv i j=1, are sampled independently from the same distribution p(D|τi). We assume that the data is given by input x label y pairs. [...] In all experiments we used an equal split of training and validation, nt i = nv i = ni. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., CPU/GPU models, memory) used for running its experiments. |
| Software Dependencies | No | The paper does not provide specific software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions). |
| Experiment Setup | No | While the paper mentions hyperparameters conceptually (e.g., learning rate, number of gradient steps), it refers to an external appendix for specific values and does not list them directly in the provided text, stating "We use a convolutional neural network commonly used with MAML on image classification (Finn, Abbeel, and Levine 2017) (see appendix of (Cioba et al. 2021))." |