[Paper Review] opML: Optimistic Machine Learning on Blockchain

K Conway, Tsz Yan So+2 more|arXiv (Cornell University)|Jan 31, 2024

Machine Learning and Data Classification4 citations

TL;DR

opML introduces an optimistic machine learning framework that enables cost-efficient, verifiable on-chain inference of large AI models using fraud proofs instead of zero-knowledge proofs. By leveraging a Turing-complete fraud-proof virtual machine, opML allows standard PCs to run 7B-parameter models like 7B-LLaMA without GPUs, achieving high accessibility and scalability while maintaining blockchain-level trust and transparency.

ABSTRACT

The integration of machine learning with blockchain technology has witnessed increasing interest, driven by the vision of decentralized, secure, and transparent AI services. In this context, we introduce opML (Optimistic Machine Learning on chain), an innovative approach that empowers blockchain systems to conduct AI model inference. opML lies a interactive fraud proof protocol, reminiscent of the optimistic rollup systems. This mechanism ensures decentralized and verifiable consensus for ML services, enhancing trust and transparency. Unlike zkML (Zero-Knowledge Machine Learning), opML offers cost-efficient and highly efficient ML services, with minimal participation requirements. Remarkably, opML enables the execution of extensive language models, such as 7B-LLaMA, on standard PCs without GPUs, significantly expanding accessibility. By combining the capabilities of blockchain and AI through opML, we embark on a transformative journey toward accessible, secure, and efficient on-chain machine learning.

Motivation & Objective

• To address the high computational cost and limited scalability of zkML for large-scale on-chain machine learning.
• To enable decentralized, verifiable, and cost-effective inference of large AI models directly on blockchain.
• To overcome the limitations of existing on-chain AI systems that rely on centralized off-chain computation or are restricted to small models.
• To support Turing-complete machine learning workloads, including model training, fine-tuning, and complex data processing, on blockchain.
• To provide a scalable, open-source alternative to zkML and non-Turing-complete systems like Agatha, enabling broader AI workloads on-chain.

Proposed method

• opML employs an optimistic rollup-style architecture, assuming all ML inference results are valid by default.
• It uses a fraud-proof virtual machine (VM) to execute ML computations off-chain and submit results on-chain.
• A challenge period allows validators to dispute incorrect results, triggering an interactive bisection protocol to pinpoint the first erroneous step.
• The dispute is resolved on-chain via minimal verification of a single computational step, ensuring low on-chain cost.
• The system supports a wide range of ML models, including deep neural networks, decision trees, KNN, and transformers, due to its Turing-complete VM.
• The framework is implemented as open-source software, enabling deployment on standard PCs without GPUs.

Figure 1 . OPML Overview

Experimental results

Research questions

• RQ1 Can large-scale machine learning models be executed on-chain efficiently and cost-effectively without relying on zero-knowledge proofs?
• RQ2 How can fraud proofs be used to achieve verifiable and decentralized on-chain inference for complex AI workloads?
• RQ3 To what extent can a Turing-complete virtual machine enable diverse machine learning tasks, including model training and data preprocessing, on blockchain?
• RQ4 Can standard hardware (e.g., CPU-only PCs) run large language models like 7B-LLaMA on-chain using an optimistic verification model?
• RQ5 How does the performance and resource efficiency of opML compare to zkML and non-Turing-complete systems like Agatha in real-world AI workloads?

Key findings

• opML enables the execution of 7B-parameters LLaMA models on standard PCs without GPUs, significantly improving accessibility.
• The system achieves cost efficiency by avoiding the high memory and computational overhead of zero-knowledge proof generation required by zkML.
• opML supports Turing-complete computation, allowing it to handle complex ML pipelines beyond simple DNN inference, such as data preprocessing and model fine-tuning.
• The fraud-proof mechanism ensures on-chain verification with minimal cost, as only a single step needs to be executed during dispute resolution.
• opML outperforms Agatha in flexibility and scope, as Agatha is limited to DNNs and lacks support for general-purpose computation and model training.
• The framework demonstrates that optimistic verification is a viable and scalable alternative to zkML for large-scale on-chain AI, especially for models that are too large for zk-proof generation.

Figure 2 . Fraud Proof VM Memory Layout