Unpacking Scattering Vision Transformer: A New Dawn in Computer Vision
summary
Introduces Scattering Vision Transformer (SVT), a vision model that replaces quadratic attention with spectral mixing (Fourier/DCT/sine bases), splits low/high frequencies and uses Einstein blending for high-frequency interactions; the architecture cuts parameters and FLOPs (small SVT ~3.9M params) and yields ~85.7% top-1 on ImageNet (≈1% gain over comparable small ViTs), trained on ImageNet-scale data and validated on downstream tasks — built, trained and faster in inference.
AI-written summary · about
chapters (11)
- 0:00Introduction to Scattering Vision Transformer
- 4:06Understanding Transformer Basics
- 9:13State Space Models vs Transformers
- 12:06Channel Mixing and Token Mixing
- 15:21Spectral Transform Techniques
- 19:29Challenges in Vision Transformers
- 22:01Scattering Vision Transformer Approach
- 25:14Model Architecture and Layers
- 28:07Data and Performance Metrics
- 30:22Comparison with Existing Models
- 32:37Applications and Adaptability
related talks
33mSeeing Through the Lens of Transformers: A New Era in Computer Vision
41mBridging Pixels and Words: The Transformer Vision Revolution
43mVisual Attention is All You Need: Introduction to Visual Transformers
17mHow Transformers Finally Ate Vision
17mVision AI in 2025
14mA Primer to Modern Day Computer Vision
55mVisual Situation Recognition
26mComputer Vision in the Era of Deep Learning