Federated Learning via Meta-Variational Dropout

Authors: Insu Jeon, Minui Hong, Junhyeog Yun, Gunhee Kim

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

Reproducibility Variable Result LLM Response
Research Type Experimental We conducted extensive experiments on various sparse and non-IID FL datasets. Meta VD demonstrated excellent classification accuracy and uncertainty calibration performance, especially for out-of-distribution (OOD) clients.
Researcher Affiliation Academia Insu Jeon Minui Hong Junhyeog Yun Gunhee Kim Seoul National University, Seoul, South Korea {insuj3on,antemrdm}@gmail.com {alsdml123,gunhee}@snu.ac.kr
Pseudocode Yes Algorithm 1 Meta VD algorithm with MAML and Reptile variant for FL
Open Source Code Yes Code is available at https://github.com/insujeon/Meta VD. ... Our code is available at https://github.com/insujeon/Meta VD.
Open Datasets Yes We used multiple FL datasets [72], including CIFAR-10, CIFAR-100, FEMINIST, and Celeb A. ... The CIFAR-10 and CIFAR-100 datasets [98] are popular for 10-class and 100-class image classification respectively. ... The Federated Extended MNIST (FEMNIST) is a widely used FL dataset for 62-class handwritten character recognition [72]. ... The Celeb A is a FL dataset based on [100] for 2-class image classification.
Dataset Splits Yes procedure LOCALADAPTATION_MAML(θ, α) ... Sample dataset Dm tr and Dm val from Dm
Hardware Specification Yes All experiments are run on a cluster of 32 NVIDIA GTX 1080 GPUs.
Software Dependencies No The paper mentions 'an optimization tool called Optuna[103]' but does not provide specific version numbers for general software dependencies like Python, PyTorch, or CUDA.
Experiment Setup Yes The batch size was set to 64, and local steps was set to 5. Personalization was executed with a batch size of 64 and a 1-step update. ... The server learning rate η was explored within the range of [0.6, 0.7, 0.8, 0.9, 1.0]. The local SGD learning rate was investigated within the range of [0.005, 0.01, 0.015, 0.02, 0.025, 0.03]. ... For Meta VD, an additional KL divergence parameter β is needed, and we sought its optimal value within the range of [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15].