Perceiver - General Perception with Iterative Attention

Explore a comprehensive analysis of Google DeepMind's Perceiver model in this 30-minute video lecture. Delve into the innovative architecture that addresses the quadratic bottleneck problem of Transformers and processes multiple input modalities simultaneously. Learn about cross-attention mechanisms, latent low-dimensional Transformers, and weight sharing across layers. Discover how the Perceiver achieves competitive performance on ImageNet and state-of-the-art results on various modalities without making architectural adjustments to input data. Gain insights into positional encodings via Fourier features, experimental results, and attention maps. Understand the potential implications of this groundbreaking research for the future of machine learning and artificial intelligence.

- Intro & Overview
- Built-In assumptions of Computer Vision Models
- The Quadratic Bottleneck of Transformers
- Cross-Attention in Transformers
- The Perceiver Model Architecture & Learned Queries
- Positional Encodings via Fourier Features
- Experimental Results & Attention Maps
- Comments & Conclusion