| Popis: |
As artificial intelligence (AI) is innovating various industries, there are concerns about the trust and transparency of AI-driven inference and training results. To tackle these issues, decentralized solutions employing blockchain have been proposed, but they often face challenges such as a long execution time of inference or training transactions, inefficient verification, and incompatibility with existing blockchains, especially when dealing with large-scale models. To overcome these limitations, we introduce BRAIN, a Blockchain-based Reliable AI Network. A unique feature of BRAIN is its two-phase transaction execution mechanism, which allows pipelined processing of inference or training transactions. Also, BRAIN performs asynchronous, aggregator-free federated learning, where each training node proposes its local model to the blockchain, computes the next global model incrementally based on the evaluation results of the proposed models, and calculates its local model using the global model and its own data, all asynchronously. For scalable verification of inference and training processes, BRAIN employs a verifiable, randomly selected committee. It reaches a consensus through a smart contract on the inference output or the evaluation score of a proposed model. Finally, BRAIN can work on an existing blockchain and run its inference and training transactions with regular transactions efficiently. Our experimental result with the GPT-J-6B model on an Ethereum-compatible chain shows that BRAIN affects the blockchain throughput gracefully, as the frequency of the inference requests reaches up to that of the most popular blockchain services today, with few timed-out requests and reasonable gas fees. This is in sharp contrast to the result of previous single-phase inference transactions. It is also shown that BRAIN converges well with asynchronous federated learning. |
