In OPML, the use of a fraud-proof VM plays a pivotal role in ensuring the correctness of ML results. This specialized VM has the capability to support general computations, encompassing both inference and training tasks. The learning process of neural networks can also be conceptualized as a series of state transitions when model parameters undergo deterministic updates.
Specifically, the inference phase of a DNN model involves straightforward forward computation on the DNN computation graph, denoted as G. On the other hand, the training process encompasses both forward computation and backward update (backpropagation) on the same DNN computation graph G. While the tasks differ in their objectives, the computation processes for forward computation and backward update are similar, enabling a unified approach.
Through the integration of a multi-phase OPML approach, we can extend support to the training process efficiently as well. Here’s how it works: The dispute protocol initiates during each iteration of the training process, enabling the identification of a dispute within a specific iteration. Subsequently, the process advances to the next phase, where the submitter and challenger engage in a dispute protocol on the computation graph for both forward and backward processes. This allows for the pinpointing of the dispute node, and the computation related to this node is then forwarded to the subsequent phase, where the fraud-proof VM arbitrates the issue.
OPML’s extension to the training process has significant implications for verifying ML model generation on the blockchain. By utilizing on-chain data to train and update the ML model, OPML ensures that the process is auditable and transparent. Moreover, OPML’s integration with the training process provides an effective means to validate ML model updates, safeguarding against potential backdoors and ensuring the model’s integrity and security.