OPML: A Machine Learning Framework Based on Optimistic Methods

We have proposed a new framework called OPML(Optimistic Machine Learning ), which can efficiently execute AI model inference and training on blockchain systems. Compared to ZKML, OPML has lower costs and higher efficiency. OPML's hardware requirements are very low, and a regular PC can run OPML tasks that include large language models such as 7B-LLaMA without a GPU.

OPML uses a verification game mechanism to ensure the decentralization and verifiability of ML services. The process is as follows:

1. The user initiates an ML service request
2. The server completes the task and submits the results to the chain.
3. Validators check the results; if there are objections, the verification game is initiated.
4. Locate specific error steps through the binary protocol
5. Finally, arbitration is carried out step by step by the smart contract.

Single-Stage Verification Game

The key points of a single-stage OPML include:

• Build a virtual machine for off-chain execution and on-chain arbitration (VM)
• Implement a dedicated lightweight DNN library to improve AI inference efficiency
• Use cross-compilation technology to compile AI inference code into VM instructions
• Use a Merkle tree to manage VM images, only upload the root hash to the chain.

In basic testing, we can complete DNN inference within 2 seconds, and the entire challenge process can be completed within 2 minutes.

Multi-Stage Verification Game

To overcome the limitations of a single-stage solution, we propose a multi-stage verification game:

• Only calculate in the VM during the final stage, other stages can be executed flexibly.
• Make full use of GPU/TPU acceleration and parallel processing
• Significantly improve execution performance, approaching local environment levels.
• Use Merkle trees to ensure the integrity and security of transitions between stages.

Taking the LLaMA model as an example, we adopt a two-stage OPML method:

• The second stage involves verification games on the computation graph, which can be accelerated using GPUs.
• The first stage converts single node computation into VM instructions.

The multi-stage method can achieve α times computational acceleration compared to the single-stage method, while significantly reducing the size of the Merkle tree.

Consistency and Determinism

To ensure the consistency of ML results, we have taken the following measures:

1. Use fixed-point algorithm ( quantization technology ) to reduce the impact of floating-point errors.
2. Use a consistent software floating-point library across platforms.

These technologies effectively overcome the challenges posed by floating-point variables and platform differences, enhancing the reliability of OPML calculations.

OPML is still under continuous development. We welcome individuals interested in this project to join and contribute to the development of OPML.