Reproduce “Social Hippocampus Memory Learning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Liping Yi, Zhiming Zhao, Qinghua Hu
SoHip (Social Hippocampus Memory Learning) is a memory-centric social machine learning framework designed for heterogeneous federated learning that exchanges abstracted memory representations among agents. This approach provides robust privacy preservation while consistently outperforming state-of-the-art baselines in accuracy by up to 8.78% on Tiny-ImageNet and achieving faster convergence under non-IID data distributions and diverse model architectures.
Key insights
- Social learning, drawing from human cognition and memory consolidation, can provide a robust paradigm for knowledge exchange in heterogeneous machine learning systems.
- Exchanging highly abstracted and compact memory representations ensures stronger privacy by avoiding direct leakage of raw data or local model parameters.
- Gated memory mechanisms, inspired by the hippocampus, allow agents to selectively integrate new short-term experiences with stable long-term knowledge for adaptive learning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.