Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].

Modular Federated Contrastive Learning with Twin Normalization for Resource-limited Clients

Authors: Azadeh Motamedi, IL MIN KIM

TMLR 2024 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Results show that Res Net18 trained with MFCL(TN) on CIFAR-10 achieves 84.1% accuracy when data is severely heterogeneous while reducing the communication burden and memory footprint compared to end-to-end training. Through experiments, we demonstrate the effectiveness of the proposed MFCL, especially with TN, which achieves robust, stable, and state-of-the-art performance on severe heterogeneous and CIB data while only a small-size client module is trained federally across clients.
Researcher Affiliation Academia Azadeh Motamedi EMAIL Department of Electrical and Computer Engineering Queen s University Il-Min Kim EMAIL Department of Electrical and Computer Engineering Queen s University
Pseudocode Yes Algorithm 1 Modular Federated Contrastive Learning (MFCL)
Open Source Code No The code will be released upon paper acceptance.
Open Datasets Yes We perform our experiments on CIFAR-10, CIFAR-100 Krizhevsky (2009), and Tiny-Image Net Le & Yang (2015).
Dataset Splits No The paper refers to standard datasets like CIFAR-10, CIFAR-100, and Tiny-Image Net, and mentions reporting accuracy on a "uniform test set". It describes how data is distributed across clients using a Dirichlet distribution but does not explicitly state the train/validation/test split percentages or sample counts for the datasets themselves within the main text, relying on standard splits of these common benchmarks.
Hardware Specification Yes We used a single NVIDIA GeForce RTX 3090 GPU to simulate the clients and r Server modules.
Software Dependencies Yes We implemented MFCL with Tensor Flow 2.14 following the standard structure of FL Mc Mahan et al. (2017) and the contrastive learning Chen et al. (2020).
Experiment Setup Yes Table 12: List of hyperparameters. Data CIFAR-10/100 Tiny-Image Net Model Res Net-18 Res Net-50 Client module (CM) First 2 layers First 4 layers CM batch size 64 64 CM optimizer Adam Adam CM learning rate 0.05 0.001 FL rounds 15 15 r Server module (r SM) epochs 150 200 r SM optimizer LARS LARS r SM learning rate cosine, lr0 = 1.0 cosine, lr0 = 1.0 Table 13: List of augmentation techniques.