MetaDiff: Meta-Learning with Conditional Diffusion for Few-Shot Learning

Authors: Baoquan Zhang, Chuyao Luo, Demin Yu, Xutao Li, Huiwei Lin, Yunming Ye, Bowen Zhang

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results show that our Meta Diff outperforms stateof-the-art gradient-based meta-learning family on FSL tasks.
Researcher Affiliation Academia 1Harbin Institute of Technology, Shenzhen 2Shenzhen Technology University baoquanzhang@hit.edu.cn, luochuyao.dalian@gmail.com, deminyu98@gmail.com, lixutao@hit.edu.cn, linhuiwei@stu.hit.edu.cn, yeyunming@hit.edu.cn, zhang bo wen@foxmail.com
Pseudocode Yes Algorithm 1: Training, Algorithm 2: Inference
Open Source Code No The paper does not provide a specific link or explicit statement indicating the availability of open-source code for the described methodology.
Open Datasets Yes Mini Imagenet. It is a subset from Image Net, which contains 100 classes and 600 images per class. Following (Lee et al. 2019), we split it into three sets, i.e., 64, 16, and 20 classes for training, validation, and test, respectively. Tiered Imagenet. It is also a Image Net subset but larger, which has 608 classes and 1200 images per class. Following (Lee et al. 2019), it is splited into 20, 6, and 8 high-level classes for training, validation, and test, respectively.
Dataset Splits Yes Mini Imagenet... we split it into three sets, i.e., 64, 16, and 20 classes for training, validation, and test, respectively. Tiered Imagenet... it is splited into 20, 6, and 8 high-level classes for training, validation, and test, respectively.
Hardware Specification No The paper does not specify the exact hardware (e.g., GPU/CPU models, memory details) used for running the experiments. It only mentions 'GPU memory' in performance analysis figures, which is a metric, not a specification of the hardware itself.
Software Dependencies No The paper mentions using 'Adam' for optimization but does not provide version numbers for any software libraries or dependencies used in the implementation.
Experiment Setup Yes Training Details. During training, we train our Meta Diff meta-optimizer 30 epochs (10000 iterations per epoch) using Adam with a learning rate of 0.0001 and a weight decay of 0.0005. Following the standard setting of diffusion models in (Ho, Jain, and Abbeel 2020), we set the number of denoising iterations to 1000 (i.e., T = 1000 is used).