LEARNING AND ADAPTATION IN COMPUTER VISION
E Telematica, Francesco Orabona · 2007
In this thesis we analyze the topics of adaptation and learning in the context of computer vision. Until now the ability of humans to adapt and learn how to solve new tasks from their own experience remain impossible to replicate in an artificial system. Even if computers can beat humans on small, constrained domains, the generality of the human mind has no counterpart in the digital world. The keys to understand and replicate a brain in a robot, could be to try to discover the general principles that govern our internal algorithms and to formalize them mathematically, and then to implement them in software. If it is true that our brains are the product of an long process of adaptation to the environment, we could be able to “predict” our biology studying the world itself. In this thesis we will show that, on one hand, it is possible to learn basic features of the processing of the neurons of the primary visual cortex from the row visual data and, on the other hand, we can learn such a high level visual skills as object classification. The obtained results support the idea that these two aspects are critical for the comprehension of biological intelligence, and, hence, for creating an artificial cognitive agent.