Ambianic.ai is an open source smart camera platform that uses on-device privacy preserving AI with helpful signals for home and workspace automation . One of the key mission objectives for Ambianic.ai is to allow users to collaboratively train novel ML models such as fall detection without sharing sensitive personal data from their homes.
Federated Learning presents an exciting opportunity towards our goal. There is explosive growth in research papers and a number of independent open source FL frameworks.
However an important challenge for FL is that of preventing bad actors from slowing down model convergence. This problem of bad actors is not unique to FL and appears in many decentralized systems without a single entity ownership. Modern blockchain systems such as Bitcoin and Ethereum have successfully addressed such problems via trustless incentive alignment that moves rational actors towards a common goal.
It is an open research question in the Ambianic.ai community whether we can introduce incentive design in the context of Ambianic.ai use cases. A few ideas are developing:
Pool users into Special Interest Groups DAO focused on collaboratively training a specific ML model (i.e. fall detection).
Require DAO members to stake tokens with tangible economic value towards the goal of the DAO.
Implement mining rewards for network nodes (actors) via Zero Knowledge Proof of Work and Validation that contributed gradients improve the overall ML model performance.
Implement slashing for bad behaving actors.
Since ML training is a stochastic rather than a deterministic process, it is not immediately obvious how to introduce a math model that effectively and efficiently measures whether a gradient snapshot contributed to an averaging round by a single node helps or harms the overall ML model performance.
Nevertheless we do have a strong conviction that web3 matters and the future of AI belongs to decentralized computing networks with strong incentive mechanisms that ensure that the network always works towards the common goals of its members. The main issue with the status quo web2 centrally trained AI systems is the asymmetric knowledge which drives incentives of service providers away from the users best interest towards shareholder best interest.
If you are interested in joining our effort to solve the collaborative privacy preserving AI training challenge in the context of home and workspace automation, check out our work and join the discussion.