Federated Contrastive Dimensionality Reduction Based on Proximal Policy Optimization

Key Points

• The aim is to develop a federated contrastive dimensionality reduction algorithm to improve visual clustering despite local data bias.
• Collaborative local model training across clients to enhance global model generalization.
• Incorporation of contrastive loss into local objective functions to correct biases from non-iid data.
• Design of an adaptive temperature adjustment agent based on proximal policy optimization.
• The proposed algorithm outperforms baseline algorithms by 16% in neighbor hit rate.
• It achieves a 9% increase in k-nearest neighbor classification accuracy.