Decentralized AI (or Decentralized ML) aims to remove central control from the lifecycle of machine learning (ML) models. From data collection to model deployment, the centralized nature of the ML lifecycle makes it tough for data science startups to compete with the large tech incumbents that have access to massive datasets, virtually unlimited computing resources and world-class research talent. This trend does not only affect startups but also big companies that can’t really compete with the Googles or Facebooks of the world in terms of AI development.
Two factors are catalyzing rapid research and growth in decentralized AI:
- Emergence of in-device ML models in mobile and internet of things (IoT) scenarios: Such models can operate without centralized server-side control.
- Increasing concern that the centralized nature of AI solutions increases the gap between large technology platforms, such as Google, Facebook, Amazon and Apple, and the rest of the world.
The research in decentralized ML is not new and can be traced back to the late 1970s with concepts like multi-agent AI systems. Decentralized ML is gaining rapid attention now because of three distinct technology trends gaining mainstream adoption simultaneously.
- Blockchains: Blockchain runtimes are a key component of decentralized ML, by modeling trustless interactions across different elements of the lifecycle of ML models such as data collection, training and optimization.
- Federated Learning: The idea that you can partition the learning process across computations in nodes within a distributed network is the hallmark of federated learning architectures and a key principle of decentralized ML.
- Private ML: Data privacy becomes really important when we remove trust in centralized authority. Techniques such as secure multiparty computations or homomorphic encryptions have been embraced in decentralized ML architectures.
Decentralized ML architectures require three main elements:
- The Runtime: Enables distributed computations without the need of a central controller. Imagine a scenario in which different parties will train a machine learning model simultaneously. The underlying runtime will be responsible for coordinating the interactions between them. In the past few years, blockchain run-times have emerged as the primary runtime for powering decentralized ML applications.
- The Code: Model representation on a decentralized network, in order to be accessed by different parties. Smart contracts in blockchain runtimes have become the preferred way to represent ML models in decentralized AI run-times. Smart contracts are like objects whose state is immutably persisted in the underlying blockchain.
- The Incentives: The absence of central coordination is replaced with incentive mechanisms that regulate the behavior of different parties. Crypto tokens are a clever mechanism for modeling incentives in decentralized ML architectures.
Further Reading:
- Harris, Justin D., and Bo Waggoner. “Decentralized and collaborative ai on blockchain.” 2019 IEEE International Conference on Blockchain (Blockchain). IEEE, 2019. Please click here to read more.
- Analysis of Models for Decentralized and Collaborative AI on Blockchain Justin D. Harris Microsoft, Montreal/Toronto, Canada Please click here to read more.
- Call themselves “the world’s first decentralized AI network” where anyone can create, share and monetize AI services at scale. Please click here to read more.
- Ocean protocol Ocean protocol provides a production-ready model for sharing datasets in a decentralized network. Please click here to read more.