Federated Learning Protocol
Last updated
Last updated
At Timeworx.io, we take data privacy, ownership and sovereignty very seriously. Federated learning is an integrated part of the decentralised data processing protocol with the target of running AI training directly on Human Agents’ smartphones.
The platform supports federated learning directly as one of the data processing types. Since the data processing will actually be carried out on the Human Agents’ smartphones, the data type that can be configured for such Tasks is an AI model. This means that whenever a Human Agent starts solving such a Task, they are actually downloading the AI model itself. By following the data processing instructions, they are training the AI model locally. And, finally, the Task outcome is nothing other than the updated AI model.
Similar to all other Tasks, when configuring the Agent block, a business is able to choose the Consensus algorithm that is actually used for aggregating the global AI model. The platform supports a wide range of Consensus algorithms that actually translate to FL model aggregation techniques:
FedAvg: one of the most commonly used methods of aggregating models in FL. During the Consensus phase, the parameters of each AI model trained by a Human Agent are weighted and averaged towards producing the global AI model.
FedProx: an enhanced version of FedAvg focused on addressing the issue of local optimisation. Running too many iterations by a Human Agent can lead to overfitting the AI model, so FedProx uses a different approach to regulate the influence of local AI models over the global model.
Scaffold: an aggregation algorithm that improves the case for data heterogeneity. Some Human Agents might be contributing data of differing degrees, and Scaffold focuses on reducing the variance of these outcomes on the global AI model.
Aside from the builtin Consensus algorithms, customers can provide their own custom implementations written in Python. The platform generates code stubs that can be implemented and deployed by the customer through a basic coding interface exposed by the UI.
We believe that federated learning is not only a privacy-enhancing technology, but it is also a key to reducing the overall environmental impact of training machine learning models. AI has become increasingly resource hungry, with reports of training processes for a single model emitting 25 times more carbon than a single air traveller flying from New York to San Francisco. In line with the Decentralised Physical Infrastructure Network (DePIN) movement, we are pushing the decentralisation bar even higher by distributing machine learning across the smartphones participating in our decentralised data processing protocol, in exchange for fair compensation.
