Scaling Few-Shot Learning for the Open World
Authors: Zhipeng Lin, Wenjing Yang, Haotian Wang, Haoang Chi, Long Lan, Ji Wang
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments show that the proposed SHA-Pipeline significantly outperforms not only the Proto Net baseline but also the state-of-the-art alternatives across different numbers of novel classes. |
| Researcher Affiliation | Collaboration | Zhipeng Lin1, Wenjing Yang1, Haotian Wang1, Haoang Chi2, 1, Long Lan1, Ji Wang1* 1 State Key Laboratory of High Performance Computing, National University of Defense Technology 2 Intelligent Game and Decision Lab, Academy of Military Science |
| Pseudocode | Yes | Algorithm 1: A Lightweight Parallel Framework, Algorithm 2: Algorithm of fine-tuning |
| Open Source Code | No | The paper does not provide any explicit statement or link for open-source code availability. |
| Open Datasets | Yes | We use the Image Net-1k((ILSVRC 2012-2017) as the base class set and construct a novel class dataset Image Net-21K-MNC based on Image Net-21K (winter 21 release), which consists of 19167 classes. |
| Dataset Splits | Yes | Among the remaining 16712 classes, 15000 classes are randomly selected for the meta-test and 1712 classes are used for the validation of meta-training. |
| Hardware Specification | Yes | All experiments are conducted on 8 NVIDIA A100 nodes with 8 GPUs each. |
| Software Dependencies | No | The paper mentions optimizers (SGD) and learning rate rules, but does not provide specific software versions for libraries like Python, PyTorch, CUDA, etc. |
| Experiment Setup | Yes | For meta-training, we use SGD optimizer without weight decay and a momentum of 0.9. The linear lr scaling rule of Goyal et al. (2017) are adopted: lr = base lr way / 5. For the learning rate schedule, we employ a cosine annealing learning rate schedule with a warm-up epoch of 5, from a base learning rate of 10-6 to 5 10-5. For Proto Net, P>M>F and Proto Net-Fix, the backbone network was meta-trained for 100 epochs, with each epoch consisting of 600 episodes. |