Toward Autonomous Intelligence: From Active 3D Vision to Invariant Object and Scene Learning, Recognition, and Search

Abstract

How do we learn what a visually seen object is? How do our brains learn without supervision to link multiple views of the same object into an invariant object category while our eyes scan a scene, even before we have a concept of the object? Indeed, why do we not link together views of different objects when there is no teacher to correct us? Why do not our eyes move around randomly? How do they explore salient features of novel objects and thereby enable us to learn view-, size-, and positionally invariant object categories? How do representations of a scene remain binocularly fused as our eyes explore it? How do we solve the Where’s Waldo problem and thereby efficiently search for desired objects in a scene? This article summarizes the ARTSCAN and ARTSCENE families of neural models, culminating in the 3D ARTSCAN Search model that clarifies how the brain solves these problems in a unified way by coordinating processes of 3D vision and figure-ground separation, spatial and object attention, object and scene category learning, predictive remapping, and eye movement search. ARTSCAN illustrates revolutionary new computational paradigms whereby the brain computes: Complementary Computing clarifies the nature of brain specialization, and Laminar Computing clarifies why all neocortical circuits exhibit a layered architecture. ARTSCAN also provides unified explanations and simulations of brain and behavioral data, and computer simulation benchmarks that support the model, which provides a blueprint for developing a new type of system for active vision and autonomous learning, recognition, search, and robotics.