Free Lunch for Few-shot Learning: Distribution Calibration

Authors: Shuo Yang, Lu Liu, Min Xu

ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In the experiments, we show that a simple logistic regression classifier trained with our strategy can achieve state-of-the-art accuracy on three datasets.
Researcher Affiliation Academia Shuo Yang1, Lu Liu2, Min Xu1 1School of Electrical and Data Engineering, University of Technology Sydney, 2Australian Artificial Intelligence Institute, University of Technology Sydney {shuo.yang, lu.liu-10}@student.uts.edu.au, min.xu@uts.edu.au
Pseudocode Yes Algorithm 1 Training procedure for an N-way-K-shot task
Open Source Code Yes The code is available at: https://github.com/Shuo Yang-1998/Few_ Shot_Distribution_Calibration
Open Datasets Yes We evaluate our distribution calibration strategy on mini Image Net (Ravi & Larochelle (2017)), tiered Image Net (Ren et al. (2018)) and CUB (Welinder et al. (2010)).
Dataset Splits Yes mini Image Net... split the dataset into 64 base classes, 16 validation classes, and 20 novel classes.
Hardware Specification No For feature extractor, we use the Wide Res Net (Zagoruyko & Komodakis, 2016) trained following previous work (Mangla et al. (2020)).
Software Dependencies No We use the LR and SVM implementation of scikit-learn (Pedregosa et al. (2011)) with the default settings.
Experiment Setup Yes Specifically, the number of generated features is 750; k = 2 and λ = 0.5. α is 0.21, 0.21 and 0.3 for mini Image Net, tiered Image Net and CUB, respectively.