Can Blockchains Reliably Train Machine Learning Models?

TL;DR

This paper proposes proof of training (PoT), a protocol that repurposes proof of work networks for reliable, secure, and scalable machine learning model training, addressing technical gaps in integration.

Contribution

Introduction of PoT protocol that aligns blockchain incentives with verifiable machine learning training, supported by theoretical analysis and a decentralized implementation.

Findings

High task throughput achieved in decentralized training network

PoT demonstrates strong robustness and security features

Potential for energy-efficient machine learning training on blockchain platforms

Abstract

Large proof of work (PoW) networks allow anyone to earn rewards by running computation-intensive hash puzzles for profit, yet they typically consume electricity comparable to that of medium-sized countries. Repurposing computing resources from hash puzzles to machine learning training can benefit the energy sector as a whole, since this computing power is no longer wasted on solving hash puzzles but is instead used to train machine learning models that provide value across different application domains. However, major technical gaps currently prevent this integration. To bridge these gaps, we introduce proof of training (PoT), a protocol that directs mining power toward verifiable training of machine learning models while preserving PoW's incentives for participation and growth. We study PoT by theoretically identifying the blockchain structure that best meets the goals of training reliability, security, and scalability, and we further evaluate it by implementing a decentralized training network. Our results indicate considerable potential, including high task throughput, strong robustness, and improved network security.