Summary of Vanishing Variance Problem in Fully Decentralized Neural-network Systems, by Yongding Tian et al.
Vanishing Variance Problem in Fully Decentralized Neural-Network Systems
by Yongding Tian, Zaid Al-Ars, Maksim Kitsak, Peter Hofstee
First submitted to arxiv on: 6 Apr 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Distributed, Parallel, and Cluster Computing (cs.DC)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary This paper explores two emerging methodologies in machine learning, Federated Learning and Gossip Learning, designed to address data privacy concerns by retaining training data on client devices and sharing locally-trained models. The primary difference between the two lies in their approach to model aggregation: Federated Learning uses a centralized parameter server, while Gossip Learning is decentralized, enabling direct model exchanges among nodes. Both methodologies involve computing a representation of received ML models and integrating this into the existing model. Our findings suggest that averaging approaches, such as FedAVG, introduce a “vanishing variance” problem, where averaging across uncorrelated models undermines optimal variance established by Xavier weight initialization. We introduce a variance-corrected model averaging algorithm that preserves optimal variance during model averaging, regardless of network topology or non-IID data distributions. Our simulation results demonstrate that this approach enables Gossip Learning to achieve convergence efficiency comparable to Federated Learning. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper is about two new ways to help computers learn from each other without sharing their personal data. “Federated Learning” and “Gossip Learning” are methods that keep training data on individual devices and only share the results with others. The main difference between them is how they combine the results: Federated Learning uses a central server, while Gossip Learning lets devices talk directly to each other. Both ways involve combining the results from different devices into one model. Researchers found that a common way of doing this, called averaging, can actually make it harder for the models to learn well. They came up with a new way to do this averaging that keeps the models learning well, even if they’re not all connected in the same way. |
Keywords
* Artificial intelligence * Federated learning * Machine learning
