DEInference

Overview

DEInference is a lightweight smart-contract built on Solana for performing decentralized inference of ML models without a dedicated L1 chain.

Features

• Privacy and ownership of your own models
• Decentralized inference with consensus based validation.
• Reputation-based rewards for model owners.
• Aggregated predictions for accurate and trustless results.

Process Flow

1. Model Registration

   • Developer trains a model off-chain.
   • Uploads a signed hash of weights and mints to immutable cNFT.
   • Reputation is set to a base value.

2. Inference Request

   • User submits an inference request specifying the input and desired model collection.
   • Smart contract broadcasts this request to model owners (or third-party computation nodes).

3. Model Execution

   • Model owners or third-party computation nodes run the models locally.
   • Predictions are returned to the smart contract.

4. Aggregation

   • Contract aggregates predictions using a weighted formula based on past model reputation.
   • Final prediction is returned to the user.

5. Reputation Management

   • Reputation scores are updated based on the prediction's accuracy relative to the aggregated result.

6. Incentivization

   • Fractions of the cost for the prediction are paid out to model owners based on their reputation scores.

How to Clone and Run (TODO)