Towards Compositional Inductive Biases and Evaluation Mechanisms for Deep Neural Networks

Abstract

The visual world is structured by human perception in terms of category information, newlinecomponent hierarchy, as well the instantiation-independent nature of several visual con- newlinecepts. All of these concepts are, again, sensitive to geometric transformations applied to newlinevisual scenes and objects. This structure plays a significant and complex role in human newlineexpectations of the visual world. In order that learning-based models for computer vi- newlinesion are both understandable in human terms and high-performing according to human newlineexpectations, these visual biases must be taken into account. Thus, models of computer newlinevision must incorporate the structure of human perception, namely the behaviour of newlinecategories under geometric transformations, the internal hierarchies of objects, and the newlineindependence of object-related concepts from their seen (during training) instances. newlineIn our work, we study certain ways in which this structure can be incorporated into vi- newlinesual learning tasks. Concretely, we develop models that are able to learn un-annotated newlinepart-whole relationships in visual objects in a transformation-robust manner. We also newlinestudy the consistency of such learnt relationships with human annotations, based on a newlineparse-tree inspired learning. Finally, we study and evaluate the instance-independence of newlineconcept-learning for visual question answering. We broadly term our subject of study as newlinequotcompositionality for human-sensibilityquot. This terminology is apt, since our work seeks newlineto incorporate part-whole compositionality and compositional learning to aid in the de- newlinevelopment of human-sensible deep neural networks. newline

Description

Keywords

Citation

item.page.endorsement

item.page.review

item.page.supplemented

item.page.referenced