Microsoft's blockchain ML experiment: training models where gas is the gradient

What it does

0xDeCA10B hosts publicly trainable ML models as Ethereum smart contracts. Anyone can query predictions for free; adding data requires passing a three-step validation through an incentive mechanism, data handler, and on-chain model update. The project includes Solidity demos with a local test chain and React dashboard, plus Python simulations to test incentive designs before deployment.

The interesting bit

The core insight is economic, not algorithmic: instead of curating data contributors, you make them stake money that they lose if their data hurts model performance. The “Deposit, Refund, and Take: Self-Assessment” mechanism essentially lets other contributors vote with their wallets to confiscate bad deposits. It’s a prediction market wearing a gradient descent costume.

Key highlights

• Supports simple on-chain models (Perceptron, Naive Bayes, Nearest Centroid) with hooks for off-chain computation via oracles or secure multiparty computation
• Three-component architecture: IncentiveMechanism (economic validation), DataHandler (on-chain storage), and model (trainable contract)
• Includes working demo with local blockchain and simulation tools for rapid mechanism testing
• Two published papers: framework overview (Blockchain-2019) and self-assessment mechanism analysis (2020)
• Explicitly acknowledges Ethereum gas costs and model complexity as fundamental constraints

Caveats

• README notes this is largely proof-of-concept; complex models require off-chain workarounds that “are not in the true spirit of this system”
• Last substantive update appears to be 2020; Ethereum has since completed its proof-of-stake transition, making some cost claims dated
• The QR code and aka.ms links suggest this was a Microsoft Research demo project, not active production infrastructure

Verdict

Worth studying if you’re designing decentralized data economies or teaching blockchain limitations through a concrete example. Skip it if you need a working, maintained ML platform—this is research code with a clever incentive mechanism at its core.