This model was realized during an intership in summer 1997 by Arnaud Delorme. It is an example of how a simple hierarchical network of integrate-and-fire neurons can detect figures in noisy images. It was the first application of SpikeNET (but it has not been published in english).

Setting up the Network

• Contrast analysis modeling the retina in the first layer
• Orientation detection in the first layer modeling neuron in V1
• 10 maps recognizing the 10 digit at every location in the image

Learning

• Supervised Learning on 5 fonts of 10 digit with random noise
• Neuron selectivity was optimized by hand

Performance of the network

Resistance to variation in size

• 4 digits with similar round shape were chosen: 0, 8, 9, 6
• During learning, the network was presented the digits at different sizes
• To test the network, a noisy image with various sizes of these digits was built
• Results show that neuron recognizing digits are very resistant to size variation