QUAN paper: real-valued quaternion convolution nets SOTA in oriented detection
Bryce Grant · hf · 2026-10-29
A new paper by Bryce Grant and Peng Wang introduces Quaternion Approximate Networks (QUAN), a framework approximating quaternion convolution with real-valued Hamilton product decompositions and custom CUDA kernels.
Key ideas
- Preserves geometric (rotation-equivariant) properties while staying efficient on real hardware
- Adds Independent Quaternion Batch Norm (IQBN) for training stability and extends quaternion ops to spatial attention
Results
- Classification (CIFAR-10/100, ImageNet): higher accuracy with fewer parameters and faster convergence vs existing convolutional and quaternion models
- Oriented detection (COCO, DOTA) and robotic perception: better parameter efficiency and rotation handling than standard CNNs, setting SOTA for quaternion CNNs
The authors highlight fit for resource-constrained robotic systems needing rotation-robust perception.
More from Embodied
- Mundane pitches a 'second body' robot that takes the shock or blast instead of tradespeople — Scobleizer · 2026-10-29
- Empatica launches WellPD, an FDA-cleared wearable for at-home Parkinson's monitoring — RosalindPicard · 2026-10-29
- Neuralink says brain implants now turn neural signals into speech at conversational speed — Polymarket · 2026-10-29
- Robotics team hosts Sim2Real playoff: win in sim, control a real robot — cixliv · 2026-10-29
- NVIDIA's ProtoMotions lets you race the same policy that drives a real Unitree G1 — xbpeng4 · 2026-10-29
- Confetti: a YC-backed platform that turns plain English into shipped hardware parts — ycombinator · 2026-10-29