Revisiting Neural Networks for Few-Shot Learning: A Zero-Cost NAS Perspective
Authors: Haidong Kang
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments show that IBFS achieves state-of-the-art performance in FSL without training, which demonstrates the effectiveness of our IBFS. |
| Researcher Affiliation | Academia | 1College of Software, Northeastern University, Shenyang, China. Correspondence to: Haidong Kang <EMAIL>. |
| Pseudocode | No | The paper describes the methodology using mathematical formulations and conceptual explanations, but it does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | Code. We have implemented our code using the Py Torch framework (Paszke et al., 2019). Specifically, for the NAS search stage, we have built upon the codebase provided by (Chen et al., 2021b). This serves as the foundation for our implementation. On the other hand, for the training and evaluation stages, we have utilized the code provided by (Tian et al., 2020). The paper mentions using existing codebases for implementation but does not explicitly state that the authors' own code for IBFS is being released or provide a link to it. |
| Open Datasets | Yes | To validate the effectiveness of the proposed proxy, we first evaluate our IBFS framework on search space of NASBench-201(Dong & Yang, 2020) in three supported datasets (CIFAR-10, CIFAR-100, Image Net-16-120 (Chrabaszcz et al., 2017)). Then, to validate the effectiveness of our IBFS in designing the FSL-friendly architecture, we conduct comprehensive experiments in two popular few-shot image classification datasets, mini-Image Net, and tiered-Image Net with the wide of peer competitors, which both are subsets of Image Net (Deng et al., 2009). |
| Dataset Splits | Yes | (1) mini-Image Net (Vinyals et al., 2016): It contains 60,000 RGB images of 84x84 pixels extracted from Image Net1K (Deng et al., 2009). It includes 100 classes (each with 600 images) that are split into 64 training classes, 16 validation classes, and 20 test classes. (2) tiered-Image Net (Ren et al., 2018): This dataset contains 779,165 RGB images of 84x84 pixels extracted from Image Net1K (Deng et al., 2009). It includes 608 classes that are split into 351 training, 97 validation, and 160 test classes. |
| Hardware Specification | Yes | Hardware. The majority of our experiments were conducted using NVIDIA RTX 2080Ti GPUs, while the remaining experiments were run on NVIDIA RTX A100 80G GPUs. Each experiment was executed on a single GPU at a time to ensure consistent and reliable results. The search cost of IBFS, specifically, was benchmarked using NVIDIA RTX 2080Ti GPUs. |
| Software Dependencies | No | Code. We have implemented our code using the Py Torch framework (Paszke et al., 2019). The paper mentions using the PyTorch framework, but it does not specify a version number for PyTorch or any other key software libraries. |
| Experiment Setup | Yes | Optimization Setup. In line with the approach proposed in (Tian et al., 2020), we employ the stochastic gradient descent (SGD) optimizer with a momentum of 0.9 and a weight decay of 0.0005. The training process for all models consists of 120 epochs for mini Image Net and 80 epochs for tiered Imagenet. Regarding the specific learning rate schedules, for mini Image Net, we start with an initial learning rate of 0.1. At epochs 40 and 80, the learning rate is decayed by a factor of 10x. As for tiered Image Net, the initial learning rate is set to 0.2. We apply a 10x learning rate decay at epochs 20, 40, and 60, 80. |