The team discusses how we should communicate our work on compositional models. We try to answer whether we should publish a paper on this topic or focus our time more on further development of Monty. We also discuss potential figures that would showcase the main advantages of learning compositional models.
Summary Video
Main Video
00:00 Intro
02:25 Benefits of Compositionality
13:23 Existing Work Looking at Compositionality: CLEVR
15:06 What New Capabilities does Compositionality Enable?
30:52 Potentially Illustrative Figures
33:41 Compositional Models Generate Rapidly and Efficiently
35:38 Compositional Models Generalize Robustly
41:56 Practical Applications from Generalization
53:14 Enabling Better Model-Based Policies with Compositionality
59:06 Visualizing Learned Compositional Models
01:04:36 Discussions
2 Likes
Hey! It’s not really clear to me from the video, are the “flat” models deep learning models? And the compositional ones monty models. Is that correct?
Hi @filipencopav , welcome to the community, and good question.
No by a flat model we mean a Monty Learning Module (LM) that has learned an object which in the real/external world is compositional in its nature (such as the mug with a logo on it), but where the LM has learned it using a single model capturing all of the object details. In other words, the representation of the logo and the mug are tangled up in the same model - as far as this LM is concerned, the mug and the logo are inseparable and always seen together, rather than different objects.
This model is “flat” in the sense that Monty has not learned to represent the object with any hierarchy (and therefore compositionality). Perhaps “non-compositional” model would be clearer.
2 Likes
