Three-Layer Optimisation Architecture
Alaya AI’s optimisation architecture consists of three layers: the Interaction Layer, the Optimisation Layer, and the Intelligent Modelling Layer. The three layers are designed to support custom data requests and provide external API access for individual users, model developers and Web3 partners with minimal privacy risks.
The three layers seamlessly integrate data collection and annotation with decentralised AI networks by combining on-chain gamification, adaptive RLHF sampling and particle swarm optimisation (PSO). A combination of hierarchal Gaussian approximation and evolutionary computation allows Alaya AI to satisfy precise AI data demands for diverse training objectives and enables HITL-assisted auto-labelling. The result is a simple yet elegant AI data solution that is highly composable and maximally compatible with both existing Web2 AI technologies and emerging Web3 AI applications.
Interaction Layer
Alaya AI’s Interaction Layer is the frontend of our platform that connects data, communities and AI technologies through a simple gamified interface, accessible through both our Google Play app and browser dApp. Users can directly access our platform through either email verification or wallet connection to contribute AI training data and earn a combination of token + NFT rewards.
Optimisation Layer
Alaya AI’s Optimisation Layer provides targeted data sampling and automated data preprocessing for optimal sampling efficiency.
The Optimisation Layer automatically verifies data quality by applying Gaussian approximation and particle swarm optimisation algorithms and enables Alaya AI to deliver superior data quality with greater efficiency and lower costs. Sampling bias is also minimised through large and diverse user communities, targeted sampling algorithms and HITL-assisted preprocessing model fine-tuning.
Intelligent Modelling Layer
Manual data labelling is too costly, inefficient, and limited in scalability for modern AI training demands. Alaya AI’s Intelligent Modelling Layer addresses this challenge by providing an infrastructure for dynamic autolabelling AI models through a combination of evolutionary computation and RLHF/HITL iteration.
Last updated