The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence) Natural Graph Networks with Taco Cohen - #440
Dec 21, 2020
Taco Cohen is a Machine Learning Researcher at Qualcomm Technologies, known for his work on equivariant networks and video compression. In this conversation, he introduces his paper on Natural Graph Networks and the concept of 'naturality,' which proposes that relaxed constraints can lead to more versatile architectures. Taco shares insights on the integration of symmetries from physics in AI, recent advances in efficient GCNNs for mobile, and innovative techniques in neural compression that significantly enhance data efficiency.
AI Snips
Chapters
Transcript
Episode notes
Equivariant Networks and Symmetry
- Taco Cohen's research focuses on equivariant networks, inspired by physics' symmetry principles.
- These networks aim to improve data efficiency in machine learning by incorporating symmetry constraints.
Natural Graph Networks
- Natural graph networks generalize equivariance, offering greater expressivity.
- They process isomorphic graphs equivalently but allow diverse computations for non-isomorphic graphs.
GCNN Demo for Medical Imaging
- Qualcomm developed a GCNN demo for NeurIPS, showcasing its application on a phone for medical imaging.
- This demonstrates efficient on-device execution of equivariant networks, addressing numerical accuracy and computational challenges.