@nleadholm gives an overview of the challenges, progress and performance metrics of compositional modeling. The presentation discussed unsupervised performance metrics such as prediction error and the challenges of Omniglot as a compositional testbed.
Summary Video
Main Video
00:00 Exploring Benchmarks for Hierarchical Learning and Compositional Object Models
08:05 When Should a Learning Module Stop Learning?
19:44 Compositional Learning at a Level in the Hierarchy
24:45 Parent and Child LMs Voting
27:56 Region 1 and Region 2 Sensing the Same Object
34:08 Unsupervised Learning Using Prediction Error
52:10 High Level Forms of Prediction Error
56:55 Attention
01:14:18 Compositional Objects and Voting
01:31:05 Omniglot Dataset
01:43:39 Compositionality in Other Written Languages
01:50:55 Short Term Takeaways for the Benchmark Test
