Context and Priming in Recursive Cortical Networks
Abstract
The recursive cortical network (RCN) is a structured probabilistic graphical model (PGM) that we have recently shown can achieve state-of-the-art performance at reading internet CAPTCHAs, a vision task at which humans excel and machines have severe difficulties. RCN provides an integrated model for two of the main problems in vision (parsing a scene into objects and recognizing the objects themselves) and solves them jointly. RCN is more data efficient than the popular deep neural networks (DNNs), which results in more robustness to unseen variations at test time.However, RCN does not tap into an important source of information for object recognition: context. The global nature of a scene might reliably predict the presence of objects in it (sceneto- object association), and similarly, some objects in a scene might predict the presence of other objects (object-to-object associations). Because DNNs optimize predictability, they already benefit to some extent from context, learning it along the actual features of an object. However, only average contexts can be learned, which lacks the flexibility and structure with which humans are able to handle context.In this project, we will modify the representational structure of RCN to make it account for context, with the goal of improving its recognition capabilities. Since RCN is a PGM (unlike DNNs), we can manipulate it to incorporate prior knowledge in a principled manner. By adding a scene variable to the model that we can condition on (and performing all the further modification that this deceivingly simple change entails), we will be able to leverage context information. We will iterate on this model, refining some of the design choices, for three particular domains of interest: CAPTCHAS, robotics and street scenes. The expected outcome is a more accurate vision system that can be informed by context, which has the potential to push the achievable performance in object recognition systems.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 23, 2019
- Source ID
- N000141912368
Entities
People
- Dileep George
Organizations
- Office of Naval Research
- United States Navy