Collaborative Learning via Prediction Consensus
Authors: Dongyang Fan, Celestine Mendler-Dünner, Martin Jaggi
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate empirically that our collaboration scheme is able to significantly boost individual models performance in the target domain from which the auxiliary data is sampled. 5 Experiments We start with a synthetic example to visualize the decision boundary achieved by our algorithm and then demonstrate its performance on real data in a heterogeneous collaborative learning setting. |
| Researcher Affiliation | Academia | 1EPFL, Switzerland 2Max Planck Institute for Intelligent Systems, Tübingen, Germany 3ELLIS Institute Tübingen, Germany |
| Pseudocode | Yes | Algorithm 1 Pseudo code of our proposed algorithm |
| Open Source Code | Yes | Code available at https://github.com/fan1dy/collaboration-consensus |
| Open Datasets | Yes | We utilize the classic Cifar10 and Cifar100 datasets [29] and A real-world dermoscopic lesion image dataset from the ISIC 2019 challenge [31, 32, 33] is included here. |
| Dataset Splits | No | The paper describes how the shared unlabeled dataset X is used for evaluation ('The evaluation metric is calculated on the dataset X') and how local data (Xi, yi) is used for training. It does not provide explicit percentages or counts for a separate validation split, nor does it refer to standard predefined train/validation/test splits for the datasets beyond stating their use (e.g., CIFAR-10). |
| Hardware Specification | Yes | All the model training was done using a single GPU (NVIDIA Tesla V100). |
| Software Dependencies | No | The paper mentions models like 'Res Net20' and 'Efficient Net' but does not specify any software dependencies with version numbers (e.g., PyTorch, TensorFlow, or specific library versions used for implementation). |
| Experiment Setup | Yes | For each local iteration, we load local data and shared unlabeled data with batch size 64 and 256 separately. The optimizer used is Adam with a learning rate 5e-3. For Cifar10 and Cifar100... do 50 global rounds with 5 local training epochs for each agent per global round. For Fed-ISIC-2019 dataset... do 20 global rounds. For the first 5 global rounds, we set λ = 0... After that, λ is fixed as 0.5. |