The visual systems of humans and higher animals exhibit a remarkable ability to recognize seen objects robustly and with high speed, tolerant of many variances for example in positional shift, view angle, illumination conditions, and unaffected by deviations from a learned prototype object or partial occlusion. Many models of the visual system assume a hierarchical set of feature detectors to play a key role for invariant object recognition. The idea is that a visual representation of a natural object is composed of a number of smaller shapes which appear more invariant under transformations than the entire object as a whole (see first image).
In this project, a large hierarchical neural network for object recognition was designed and implemented on a neuromorphic hardware sytstem. The feature detectors in the hidden layers were trained by a novel self-organization algorithm requiring only very few training data consisting of images of the objects to be recognized. Good classification results are obtained on the MNIST benchmark of hand-written digits and on traffic sign image
Keywords: Convolutional Neural Networks; Self-Organzation; Competitive Learning; Analog Neuromorphic Hardware; Image Recognition
Johannes Fieres: A Method for Image Classification Using Low-Precision Analog Computing Arrays. Dissertation, University of Heidelberg (2006) Download (1.7 MB)
J. Fieres, J. Schemmel, K. Meier: Training convolutional neural networks of threshold neurons suited for low-power hardware implementation. IJCNN2006, 21--28, IEEE Press (2006) Details/pdf
J. Fieres, J. Schemmel, K. Meier: A convolutional neural network tolerant of synaptic faults for low-power analog hardware. ANNPR 2006, Springer Lecture Notes in Artificial Intelligence 4087, 122-132 (2006) Details/pdf
A biologically inspired, hierarchical network model for object recognition.
Images used for testing the system: Traffic signs and hand-written digits (MNIST data base).
Real-time image processing demonstration using the analog hardware system.