Humans can recognize objects in a cluttered scene in 100s of milliseconds. Computer algorithms operate at a much lower performance level compared to humans. Furthermore it has proven to be particularly difficult to recognize all objects in a category, such as, all cat faces vs dog faces instead of a particular cat or dog face. In this example there is a large in-class variability. The distinguishing features can vary significantly among different objects in the same class. This variation of features within one class is a difficult problem to address. A similar case can be made for other categories, such as, cars, human faces, etc.
The human visual system can be divided into two major pathways, commonly called the 'what' and 'where' pathways. The 'what' pathway recognizes an object in a scene, but not its specific location. In this presentation I describe a biologically inspired hierarchical 'what' neural network that can successfully classify objects into categories.
Colloquia Series page.