Human-level concept learning through probabilistic program induction
凭借概率规划归纳法进行人类层级的概念学习
Summary:
As we all know that machine learning algrithms tend to performe better with more trainning examples. While people can learn a concept from a example or few examples. On one-shot classification task，this paper shows us a Bayesian program learning(BPL) framework, BPL learns concepts like human do，it can get the essence of handwritten characters with few examples. It incorporates the principles of compositionality, causality, and learning to learn. Isn't it a good idea of combining DCNN with BPL ?
image.png
