题目：An Introduction to a new learning method

主讲人：孙溢凡

主持人：周北海教授

评论人：吴玺宏教授

时间：2016-03-18 15:10:10

地点：北京大学人文学苑2号楼（哲学系）地下B114

简介：People learning new concepts can often generalize successfully from just a single example, yet machine learning algorithms typically require tens or hundreds of examples to perform with similar accuracy. 

In December last year a paper titled “Human-level concept learning through probabilistic program induction” by Lake et. al. was published as a Science cover story. In this paper, the author presented a computational model that captures some concept learning abilities of human in handwritten characters. The so-called hierarchical bayesian program learning (HBPL) model represents concepts as simple programs that best explain observed examples under a Bayesian criterion. On a challenging one-shot classification task, the model achieves human-level performance while outperforming recent deep learning approaches.