A model compression algorithm with edge-end fusion enhancement that leverages “crowd-sourced collaborative computing and end-adaptive learning” methods to adaptively blueuce model computation and achieve efficient, accelerated computation.
A runtime dynamic self-evolving model compression framework, which is most suitable for selecting model compression strategies and deploying deep models in the current context,achieving optimal utilization of IoT terminal resources.
Adadeep integrates the adaptive model compression problem with a DNN hyperparameter optimization framework, using reinforcement learning to achieve adaptive lightweight model architecture search.
A lightweight near-real-time low-light video enhancement middleware，It is capable of enhancing low-light video streams in near-real-time using only CPU resources on edge devices.
A Context-aware adaptive quantization framework that dynamically switches different gates based on the resource environment of the model to generate mixed-precision quantization strategies that match the hierarchical structure of the backbone network.
The rapid adaptive search problem for identifying model breakpoints in a dynamic collaborative group consisting of multiple coexisting intelligent agents in GADS algorithm exploration.
Cross-FCL explores the enhancement of continuous learning capabilities for cross-edge devices through the method of “parameter separation + cross-edge knowledge fusion” to balance memory and adaptability.
MetaDetector is a fast method for constructing specific fake news detection models in scenarios where there is a scarcity of high-quality training data for newly occurring events and sparse label information.
The AskMe algorithm utilizes techniques such as multi-view, Bi-LSTM, and attention to effectively integrate the imbalanced multi-behavior of users and similar behavior of group users, guiding efficient detection of rare behaviors.