Meta Discovery: Learning to Discover Novel Classes given Very Limited Data

Authors: Haoang Chi, Feng Liu, Wenjing Yang, Long Lan, Tongliang Liu, Bo Han, Gang Niu, Mingyuan Zhou, Masashi Sugiyama

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental This meta-learning-based methodology significantly reduces the amount of unlabeled data needed for training and makes it more practical, as demonstrated in experiments.
Researcher Affiliation Academia 1National University of Defense Technology 2Hong Kong Baptist University 3University of Technology Sydney 4The University of Sydney 5RIKEN AIP 6The University of Texas at Austin 7The University of Tokyo
Pseudocode Yes Algorithm 1 Clustering-rule-aware task sampler (CATA); Algorithm 2 MM for NCDL.; Algorithm 3 MP for NCDL.
Open Source Code Yes To ensure the reproducibility of experimental results, we have provided codes of MM and MP at github.com/Haoang97/MEDI.
Open Datasets Yes We conduct experiments on four popular image classification benchmarks, including CIFAR-10 (Krizhevsky & Hinton, 2009), CIFAR-100 (Krizhevsky & Hinton, 2009), SVHN (Netzer et al., 2011), and Omni Glot (Lake et al., 2015).
Dataset Splits No The paper defines training and test sets for individual tasks within the meta-learning framework (Sl,tr i and Sl,ts i). However, it does not specify a separate, global validation dataset split for the entire experimental setup.
Hardware Specification Yes We implement all methods by Py Torch 1.7.1 and Python 3.7.6, and conduct all the experiments on two NVIDIA RTX 3090 GPUs.
Software Dependencies Yes We implement all methods by Py Torch 1.7.1 and Python 3.7.6, and conduct all the experiments on two NVIDIA RTX 3090 GPUs.
Experiment Setup Yes The output dimension of feature extractor πmm is set to dr = 512. The meta learning rate and inner learning are 0.4 and 0.001 respectively. We use a meta batch size (the amount of training tasks per training step) of 16\8 for {CIFAR-10,SVHN}\{CIFAR-100,Omniglot}. In addition, we choose k to be 10 which is suitable for all datasets. For each training task, we update the corresponding inner-algorithm by 10 steps.