Decentralized AI — 0xindexsan
Public.icon
As long as the data is public, no matter how accurate and private the learning model is, it will not provide competitive advantage. Rather, the fact that the learning model is public, along with its architecture, incentive design, and development community, will bring about an advantage in the future.
So let's consider the three conditions for making AI public. I think the necessary conditions are as follows:
Democratization of data: Data is held by blockchain or end users, not central servers
Decentralization of machine learning: Learning/inference of AI is executed by distributed servers, not central servers
Public access to the model: No restrictions on access to the AI model
Making AI public increases the risk of information leakage and exposure to malicious attacks. This is a similar issue to the 51% attack in blockchain. Technologies to reduce the risk of exposure to malicious attacks?
This seems to be the most useful?
Federated Learning is one of the most promising machine learning architectures in the field of distributed AI.
With Federated Learning, multiple participants of AI nodes can train or optimize AI models individually without trusting each other or a central authority.
The rules of Federated Learning are as follows:
End users have ML models on their own devices and run learning at the time their data is "born"
Share the learning results with the whole
Update the overall model
https://gyazo.com/e259f6ca0c136c5bb4ae7a25bd00a7f8
Related projects include
https://gyazo.com/f641da8cd11c9265ad41a2983722e1f0