Generative factorization for object-centric representation learning

Abstract

Empowering machines to understand compositionality is considered by many (Lake et al., 2017; Lake and Baroni, 2018; Schölkopf et al., 2021) a promising path towards improved representational interpretability and out-of-distribution generalization. Yet, discovering the compositional structures of raw sensory data requires solving a factorization problem, i.e. decomposing the unstructured observations into modular components. Handling the factorization problem presents numerous technical challenges, especially in unsupervised settings which we explore to avoid the heavy burden of human annotation. In this thesis, we approach the factorization problem from a generative perspective. Specifically, we develop unsupervised machine learning models to recover the compositional data-generation mechanisms around objects from visual scene observations.

First, we present MulMON as the first feasible unsupervised solution to the multi-view object-centric representation learning problem. MulMON resolves the spatial ambiguities arising from single-image observations of static scenes, e.g. optical illusions and occlusion, with a multi-view inference design. We demonstrate that not only can MulMON perform better scene object factorization with less uncertainty than single-view methods, but it can also predict a scene's appearances and object segmentations for novel viewpoints. Next, we present a technique, namely for latent duplicate suppression (abbr. LDS), and demonstrate its effectiveness in fixing a common scene object factorization issue that exists in various unsupervised object-centric learning models---i.e. inferring duplicate representations for the same objects. Finally, we present DyMON as the first unsupervised learner that can recover object-centric compositional generative mechanism from moving-view-dynamic-scene observational data. We demonstrate that not only can DyMON factorize dynamic scenes in terms of objects, but it can also factorize the entangled effects of observer motions and object dynamics that function independently. Furthermore, we demonstrate that DyMON can predict a scene's appearances and segmentations at arbitrary times (querying across time) and from arbitrary viewpoints (querying across space)---i.e. answer counterfactual questions.

The scene modeling explored in this thesis is a proof of concept, which we hope will inspire: 1) a broader range of downstream applications (e.g. "world modelling'' and environment interactions) and 2) generative factorization research that targets more complex compositional structures (e.g. complex textures, multi-granularity compositions).