@nleadholm outlines our current thinking on compositionality and how to measure performance when multiple columns sense the same object while recognizing its subcomponents. The team then discusses candidate datasets for benchmarking compositional learning and evaluation.
Watch an introduction of this video here:
00:00 Introduction
01:13 Testing for Compositionality
05:33 How to Evaluate Monty Performance with Multiple Columns
43:30 Omniglot Dataset
59:09 TBP Generated Data Set
01:11:46 Compositional Evaluation
