Disentangling What and Where for 3D Object-Centric Representations Through Active Inference

Abstract

Although modern object detection and classification models achieve high accuracy, these are typically constrained in advance on a fixed train set and are therefore not flexible to deal with novel, unseen object categories. Moreover, these models most often operate on a single frame, which may yield incorrect classifications in case of ambiguous viewpoints. In this paper, we propose an active inference agent that actively gathers evidence for object classifications, and can learn novel object categories over time. Drawing inspiration from the human brain, we build object-centric generative models composed of two information streams, a what-and a where-stream. The what-stream predicts whether the observed object belongs to a specific category, while the where-stream is responsible for representing the object in its internal 3D reference frame. We show that our agent (i) is able to learn representations for many object categories in an unsupervised way, (ii) achieves state-of-theart classification accuracies, actively resolving ambiguity when required and (iii) identifies novel object categories. Furthermore, we validate our system in an end-to-end fashion where the agent is able to search for an object at a given pose from a pixel-based rendering. We believe that this is a first step towards building modular, intelligent systems that can be used for a wide range of tasks involving three dimensional objects.

Designing ecosystems of intelligence from first principles

“…Relatedly, active inference provides a promising framework for learning representations in which distinct generative factors are disentangled (Higgins et al, 2018), via the sensorimotor contingencies associated with controllable latent factors (Hinton et al, 2011; Tschantz et al, 2020; Van de Maele et al, 2021). Low-dimensional disentangled representations, in addition to being useful for an AI system itself in achieving its own ends, are more explainable and human-interpretable than generic latent representations.…”

Embodied Object Representation Learning and Recognition

“…Building on previous work (Van de Maele et al, 2021a ), we now evaluate our agent on pixel data rendered from 33 objects from the YCB benchmarking dataset (Calli et al, 2015 ). In this article, we show that using object-specific models introduces the ability to classify out-of-distribution objects through a two-stage process that first aggregates the votes and then compares the prediction error on the likelihood of the observation.…”