Research Internships: Feedforward Network Training for Object Recognition
Code of project: 2009-03
Project Supervisors: Dr Marc de Kamps
Area of interest: This project is linked to the Biosystems Research Group.
Appropriate for degree programmes: All single and joint honours students from the School of Computing
Pre-requisites:The candidate must have considerable programming experience. Knowledge of C++ is a plus, but much of the project can be programmed in Python. Further information about the prerequisites for this project is available from Dr Marc de Kamps.
Project Outline
Biological creatures still outperform machines in object recognition and classification. As more becomes known about primate vision, researchers are trying to create ever more realistic models of the neural networks that underlie these object recognition capabilities. This has led to a number of models of the ventral stream in cortex, which is a network implicated in object recognition.
Artificial models of the ventral stream have shown good performance in object recognition for isolated objects, but perform poorly in multi-object scenes. From neuroscience it is known that feedback activity plays an important role in keeping the complexity of image processing under control. This feedback activity allows prior knowledge or expectations which are represented in high level cortical areas to interact with visual information in lower layers of the cortex and help to process visual information selectively; only specific object features or specific locations are processed.
So far, most biological models of the ventral stream have not employed feedback activity. The goal of this project is to extend an existing feedforward model of the ventral stream to include feedback connections. As a first test of such a feedback architecture, it will be attempted to locate a familiar object in a cluttered visual scene.
The candidate for this project will work from an existing implementation of the feedforward network and will train the network to recognize simple objects. Then, multi-object scenes will be created and presented to the network. The performance of the network for such multi-object scenes will be evaluated. Then one object will be designated the target object and the feedback network will be used in order to locate the target object in lower visual areas among the other objects.
A successful implementation of this mechanism would be a demonstration of object-based attention in a biological model of vision, which would be new.