OPML: An Optimistic Approach to Machine Learning Systems

This article introduces a new blockchain system called OPML( Optimistic Machine Learning ), which utilizes an optimistic approach for AI model inference and training/fine-tuning. Compared to ZKML, OPML can provide more cost-effective ML services.

One of the major advantages of OPML is its low barrier to entry for participation. Currently, ordinary PCs can run OPML systems that include large language models like the 26GB 7B-LLaMA( without the need for a GPU. This system employs a verification game mechanism to ensure the decentralization of ML services and verifiable consensus.

The process to verify the game is as follows:

1. The requester initiates the ML service task.
2. The server completes the task and submits the result to the chain.
3. The validator verifies the results, and if there are any disputes, a verification game is initiated.
4. Finally, conduct step-by-step arbitration on the smart contract.

![OPML: Machine Learning Using Optimistic Rollup System])https://img-cdn.gateio.im/webp-social/moments-e798407b4f5f3dd6dc7d8327db07eb20.webp(

OPML adopts two verification game modes: single-stage and multi-stage. The single-stage mode is similar to calculating delegation )RDoC(, by precisely pinpointing the disputed steps and submitting them for arbitration by on-chain contracts. To improve efficiency, OPML has also developed a dedicated lightweight DNN library and virtual machine system.

![OPML: Machine Learning Using Optimistic Rollup System])https://img-cdn.gateio.im/webp-social/moments-3f290bc2a1acee40d4e71fbf35ee5079.webp(

Multi-stage verification games overcome the limitations of single-stage models by fully utilizing GPU/TPU acceleration and parallel processing capabilities. They gradually narrow down the scope of disputes through multiple stages, ultimately pinpointing specific VM instructions. This approach significantly improves the execution efficiency of OPML, bringing its performance close to that of a local environment.

![OPML: Machine Learning Using Optimistic Rollup System])https://img-cdn.gateio.im/webp-social/moments-4d41ed09832980b943519f4c0baa6109.webp(

To ensure cross-platform consistency, OPML adopts fixed-point algorithms and software floating-point libraries. Compared to ZKML, OPML has significant advantages in terms of computational efficiency, flexibility, and universality, providing new possibilities for decentralized AI applications.

![OPML: Machine Learning Using Optimistic Rollup System])https://img-cdn.gateio.im/webp-social/moments-a33f120074b07b2ec4ae4ececbea79f1.webp(

The OPML project is still actively in development, and interested developers are welcome to participate and contribute.