AI Capsule System Distinguishes Objects with State-of-the-Art Precision

2019-12-04

Strong evidence suggests that humans always rely on coordinate systems or reference lines and curves to infer the position of points in space. This is different from widely used computer vision algorithms, which often distinguish objects through the numerical representation of object features. In order to seek a more human-like approach for machines, researchers from Google, Alphabet subsidiary DeepMind, and the University of Oxford have proposed a Stacked Capsule Autoencoder (SCAE) that uses the geometric relationships between different parts of an object to reason about the object. Since these relationships do not depend on the model's view of the object's position, the model can accurately classify the target even when the view changes. In 2017, Geoffrey Hinton, a leading theorist and Turing Award winner in the field of artificial intelligence, along with students Sara Sabour and Nicholas Frost, introduced a machine learning architecture called "CapsNet," which is a differentiable multi-level approach that achieves state-of-the-art image classification performance on popular benchmarks. Earlier this year, Hinton, Sabour, and researchers from the Oxford Robotics Institute provided a detailed description of SCAE, which improves upon the original architecture in several key aspects. SCAE, like other capsule systems, interprets organized collections of parts through geometry to understand objects. Specifically, a set of mathematical functions called capsules, which are responsible for analyzing various object attributes such as position, size, and hue, are added to an artificial intelligence model commonly used for analyzing visual images. The predictions of multiple capsules are then reused to form a representation of the parts. Since these representations are preserved throughout the analysis process of SCAE, the capsule system can utilize them to identify objects, even when parts are swapped or transformed. While capsule systems are highly unique, their functionality, like all deep neural networks, is organized in interconnected layers. The difference lies in the dynamic computation of the weights of capsules, which are based on the ability of the previous layer's functions to predict the next layer's output, rather than data transmission signals. SCAE consists of three stages. In the first stage, the Constellation Capsule Autoencoder (CCAE) is used to extract the pixels of the image to be analyzed. In the second stage, the Part Capsule Autoencoder (PCAE) segments the image into constituent parts and infers their poses before reconstructing the image. In the third stage, the Object Capsule Autoencoder (OCAE) attempts to integrate the discovered parts and their poses into a smaller set of objects and then tries to reconstruct them.

RECENT AI NEWS

RECENT AI TOOLS