NIPS読み会2013: One-shot learning by inverting a compositional causal processnozyh
This document summarizes research on one-shot learning using a hierarchical Bayesian program learning (HBPL) model. The key points are:
- The HBPL model achieved human-level performance on one-shot classification and generation tasks, outperforming other deep learning models.
- It used an Omniglot dataset of handwritten characters from different alphabets to learn concepts from a single example, as well as produce new examples.
- The model learns "motor programs" that represent common patterns in how symbols are drawn, from a library of primitive strokes. This allows it to generalize concepts from limited data.