Learning a Few-shot Embedding Model with Contrastive Learning
Authors: Chen Liu, Yanwei Fu, Chengming Xu, Siqian Yang, Jilin Li, Chengjie Wang, Li Zhang8635-8643
AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our model is thoroughly evaluated on few-shot recognition task; and demonstrates state-of-the-art results on mini Image Net and appealing performance on tiered Image Net, Fewshot-CIFAR100 (FC-100). To validate our method, we conduct experiments on several widely used datasets. Table 1: 5-way few-shot accuracies with 95% confidence interval on mini Image Net and tiered Image Net. Ablation Study: For our method, we have different parts:info NCE, hard sample, Patch Mix. As shown in Tab. 5, each part contributes to the improvement. |
| Researcher Affiliation | Collaboration | Chen Liu* 1, Yanwei Fu* 1, Chengming Xu 1 , Siqian Yang2, Jilin Li 2 , Chengjie Wang 2 , Li Zhang 1 1 School of Data Science, and MOE Frontiers Center for Brain Science, Fudan University 2 You Tu Lab Tencent |
| Pseudocode | No | The paper describes algorithms but does not provide a structured pseudocode block or algorithm listing. |
| Open Source Code | No | The paper does not provide an explicit statement or link for open-source code for the methodology described. |
| Open Datasets | Yes | mini Image Net (Vinyals et al. 2016) is a sub-dataset from Image Net (Russakovsky et al. 2015). tiered Image Net is also sampled from Image Net (Russakovsky et al. 2015)... Fewshot-CIFAR100 (FC-100) dataset (Oreshkin, Rodriguez, and Lacoste 2018) is a subset of CIFAR-100. |
| Dataset Splits | Yes | mini Image Net...These categories are split into train, val and test with 64, 16 and 20 classes respectively. The partition follows the instruction of (Ravi and Larochelle 2017). tiered Image Net...separated into 351 classes for training, 97 for validation and 160 for testing as suggested in (Ren et al. 2018). Fewshot-CIFAR100 (FC100)...A common split is 60, 20 and 20 categories for train, val and test set. |
| Hardware Specification | No | The paper does not specify any hardware details like GPU or CPU models used for running experiments. |
| Software Dependencies | No | The paper does not specify any software dependencies with version numbers. |
| Experiment Setup | Yes | Implementation Details. Res Net12 is our selected model structure, the details follow the one proposed in TADAM (Oreshkin, Rodriguez, and Lacoste 2018). We use he-normal (He et al. 2015) to initialize the model. Stochastic Gradient Descent(SGD) (Bottou 2010) is taken as our optimizer. The initial learning rate is 0.1. For mini Image Net, we decrease the learning rate at 12, 000-th, 14, 000-th and 16, 000-th episode. For tiered Image Net, the learning rate is halved at every 24,000 episodes. For all the experiments, we test the model for 2000 episodes. 4 episodes are picked for every batch during training. Images of tiered Image Net and mini Image Net are firstly resized to 84 84 during training and testing process. Images of FC100 are resized to 32 32. For training process, random horizontal flip and random crop are utilized as common data augmentation as used in (Hou et al. 2019). ... We set the weights for supervised loss and info NCE loss as 1 and 0.5. ... We choose three kinds of grid size 1 1, 6 6 and 11 11. |