Chimera Framework Enables Trustworthy Neuro-Symbolic AI on Programmable Dataplanes
A new study has introduced Chimera, a framework designed to implement complex learning models right onto programmable dataplanes for better and quicker traffic analysis. It addresses challenges posed by strict hardware constraints and the need for network behavior that is both predictable and auditable. Chimera combines attention-driven neural computations with symbolic constraints within dataplane primitives, allowing for dependable inference in the match-action pipeline. It incorporates a kernelized attention approximation and a two-layer key-selection structure along with a cascade fusion technique to ensure symbolic guarantees while preserving neural capabilities. Additionally, the framework features a hardware-aware mapping protocol and a dual update system, guaranteeing stable performance under real-world dataplane conditions. The research, showcasing the power of neuro-symbolic approaches, is available on arXiv with the identifier 2602.12851v3.
Key facts
- Chimera is a framework for deploying learning models on programmable dataplanes
- It enables line-rate, low-latency traffic analysis
- The framework maps attention-oriented neural computations onto dataplane primitives
- It combines kernelized attention approximation with symbolic constraints
- A two-layer key-selection hierarchy is part of the architecture
- Cascade fusion mechanism enforces hard symbolic guarantees
- Hardware-aware mapping protocol and two-timescale update scheme enable stable operation
- Research was published on arXiv with identifier 2602.12851v3
Entities
Institutions
- arXiv