Presentation
QuadraNet: Improving High-Order Neural Interaction Efficiency with Hardware-Aware Quadratic Neural Networks
DescriptionRecent progress in Computer Vision is mostly driven by high-order interaction. However, current high-order interactive neural networks suffer from computation bottlenecks for large intermediate state space or data dependency. To address this challenge, we propose a neural-level approach, namely QuadraNet, using the quadratic neuron and an effective and efficient network design. QuadraNet becomes hardware-friendly by optimizing the intermediate state space and eliminating recursive design, while it only uses a well-optimized operator library. To achieve maximum QuadraNet's hardware-friendliness, we combine it with Hardware-aware Neural Architecture Search. Experiments show that QuadraNet achieves a speedup of high-order interactive networks while maintaining high performance.
Event Type
Work-in-Progress Poster
TimeTuesday, July 11th6:00pm - 7:00pm PDT
LocationLevel 2 Lobby
AI
Autonomous Systems
Cloud
Design
EDA
Embedded Systems
RISC-V
Security