RankDNN: Learning to Rank for Few-Shot Learning

Authors: Qianyu Guo, Gong Haotong, Xujun Wei, Yanwei Fu, Yizhou Yu, Wenqiang Zhang, Weifeng Ge

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments demonstrate that Rank DNN can effectively improve the performance of its baselines based on a variety of backbones and it outperforms previous state-of-the-art algorithms on multiple few-shot learning benchmarks, including mini Image Net, tiered Image Net, Caltech-UCSD Birds, and CIFAR-FS. Furthermore, experiments on the cross-domain challenge demonstrate the superior transferability of Rank DNN.
Researcher Affiliation Academia Qianyu Guo1,2, Gong Haotong1, Xujun Wei1,3, Yanwei Fu2, Yizhou Yu4, Wenqiang Zhang2.3, Weifeng Ge1,2* 1Nebula AI Group, School of Computer Science, Fudan University,Shanghai,China 2Shanghai Key Laboratory of Intelligent Information Processing,Shanghai,China 3Academy for Engineering & Technology, Fudan University,Shanghai,China 4Department of Computer Science, The University of Hong Kong,Hong Kong,China wfge@fudan.edu.cn
Pseudocode Yes Algorithm 1: Meta-Training of Rank DNN for Few-shot Learning
Open Source Code Yes The code is available at: https://github.com/guoqianyu-alberta/Rank DNN.
Open Datasets Yes We use four popular benchmark datasets in our experiments: mini Image Net(Vinyals et al. 2016), tiered Image Net (Ren et al. 2018), Caltech-UCSD Birds-200-2011 (CUB)(Chen et al. 2019b), and CIFAR-FS(Bertinetto et al. 2018).
Dataset Splits Yes All datasets follow a standard division and all images are resized to predefined resslutions following standard settings. We evaluate the performance under standard 5-way-1-shot and 5-way-5-shot settings.
Hardware Specification No No specific hardware details (such as CPU, GPU models, or memory) used for running the experiments are provided in the paper.
Software Dependencies No The paper does not provide specific software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions or specific library versions).
Experiment Setup Yes For the feature extractor, we consider three state-of-the-art backbones, S2M2 (Mangla et al. 2020), IE (Rizve et al. 2021) and FEAT (Ye et al. 2020). The number of neurons in different layers of Rank MLP is [6400, 1024, 512, 256, 1] for S2M2 and FEAT, and [16384, 1024, 512, 256, 1] for IE. We set the weight decay to 10^-6 and the momenta to 0.9. The learning rate is fixed at 0.0005 for both networks. Note that during the meta test stage, Rank DNN does not need to finetune on 1-shot, but on 5-shot, Rank MLP needs to be finetuned with the support set to get good performance, where we sample 100 triplets randomly in each mini-batch, and the parameters of Rank MLP is updated in 100 iterations. The learning rate is set to 0.01. Rank DNN is optimized with SGD.