PADO: A New learning Architecture For Object Recognition
Astro Teller, Manuela Veloso · 1997
Abstract In general, AI systems use symbols to represent knowledge and to reason about tasks. Most of these systems today still work on simple problems and artificial domains. One of the main reasons for this is the common assumption that sensing is not only perfect, but also that sensors return specific symbols, not raw data. The signal-to-symbol problem is the task of converting raw sensor data into a set of symbols that the data can be seen as representing. One of the main goals of computer vision is to provide a solution to the signal to-symbol problem. In particular, this goal involves object recognition, that is, the ability to recognize what (and where) objects are shown in an image. Machine learning can do induction on a set of examples to learn to discriminate among classes. These two fields, machine learning and computer vision, are natural mates and are particularly suited to cooperate on object recognition tasks. Several proven machine learning architectures, such as neural networks, have been integrated with computer vision, and the results in the recognition of “everyday” objects have been modest at best. It is possible that better parameter values, more training data, or faster computers will allow one of these architectures to make some significant advance in the field of object recognition. This chapter proposes a different view: given the experienced difficulties with the current architectures, a more profitable path is to investigate new architectures.