Explore the innovative HyperTransformer model for few-shot learning in this comprehensive video featuring a paper explanation and author interview. Dive into the architecture that generates convolutional neural network weights directly from support samples, decoupling task space complexity from individual task complexity. Learn about weight generation versus fine-tuning, the benefits of self-attention mechanisms, and the potential for producing larger weights. Analyze experimental results, discuss open questions, and gain insights into the model's effectiveness for small target CNN architectures and semi-supervised learning scenarios.
HyperTransformer - Model Generation for Supervised and Semi-Supervised Few-Shot Learning
Overview
Syllabus
- Intro & Overview
- Weight-generation vs Fine-tuning for few-shot learning
- HyperTransformer model architecture overview
- Why the self-attention mechanism is useful here
- Start of Interview
- Can neural networks even produce weights of other networks?
- How complex does the computational graph get?
- Why are transformers particularly good here?
- What can the attention maps tell us about the algorithm?
- How could we produce larger weights?
- Diving into experimental results
- What questions remain open?
Taught by
Yannic Kilcher
Related Courses
-
AI skills for Engineers: Supervised Machine Learning
-
Big Self-Supervised Models Are Strong Semi-Supervised Learners
-
Neural Fine-Tuning Search for Few-Shot Learning
-
Flexible Neural Networks and the Frontiers of Meta-Learning
-
Towards Generalist Imaging Using Multimodal Self-Supervised Learning - Mars Huang
