Synthesizer - Rethinking Self-Attention in Transformer Models
Offered By: Yannic Kilcher via YouTube
Course Description
Overview
Dive into a comprehensive video analysis of the research paper "Synthesizer: Rethinking Self-Attention in Transformer Models". Explore the revolutionary concept of synthetic attention weights in Transformer models, challenging the necessity of dot-product attention. Learn about Dense Synthetic Attention, Random Synthetic Attention, and their comparisons to traditional feed-forward layers. Examine experimental results across various natural language processing tasks, including machine translation, language modeling, summarization, dialogue generation, and language understanding. Gain insights into the performance of the proposed Synthesizer model against vanilla Transformers, and understand the implications for future developments in attention mechanisms and Transformer architectures.
Syllabus
- Intro & High Level Overview
- Abstract
- Attention Mechanism as Information Routing
- Dot Product Attention
- Dense Synthetic Attention
- Random Synthetic Attention
- Comparison to Feed-Forward Layers
- Factorization & Mixtures
- Number of Parameters
- Machine Translation & Language Modeling Experiments
- Summarization & Dialogue Generation Experiments
- GLUE & SuperGLUE Experiments
- Weight Sizes & Number of Head Ablations
- Conclusion
Taught by
Yannic Kilcher
Related Courses
Axial-DeepLab - Stand-Alone Axial-Attention for Panoptic SegmentationYannic Kilcher via YouTube Linformer - Self-Attention with Linear Complexity
Yannic Kilcher via YouTube The Narrated Transformer Language Model
Jay Alammar via YouTube Learning the Structure of EHR with Graph Convolutional Transformer - Edward Choi
Stanford University via YouTube Knowledge Is Embedded in Language Neural Networks but Can They Reason?
Simons Institute via YouTube