Professor Mohammad Ali Maddah-Ali at ECE Spring 2024 Colloquium

Blockchains, decentralized learning, and game of coding

The decentralization movement, which emerged in 2008 with the success of Bitcoin, aims to revolutionize digital platforms by making them transparent and open to the public for coordination, contribution, and verification. This movement is currently reshaping the AI landscape, with a focus on designing decentralized machine learning (DeML) platforms. However, blockchains, as decentralized trust engines, face limitations in computational power, which hinder this transformation.

In this talk, we will first review blockchain consensus, its potential to enable distributed computing platforms, and its inherent limitations. We will then discuss why existing solutions based on verifiable computation outsourcing are ineffective for massive and approximate ML computations. As an alternative, we will introduce the game of coding, which leverages the power of redundant (coded) computing to effectively detect and correct errors in outsourced tasks. The game of coding utilizes the inherent rationality of potential adversaries to expand the applicability of coding theory in a trust-minimized setting, where external nodes are overwhelmingly dishonest.