Precursor networks for training the binary perceptron

摘要

Considering the storage problem for the binary perceptron Penney and Sherrington studied the properties of clipping the maximum stable continuous network and calculated the fraction of weights that correctly predict the binary weights. One can wonder whether there are better continuous networks than the maximum stable one to start from. In this work other starting vectors (precursors) will be presented which perform better on clipping than the maximum stable perceptron. We will show that precursors obtained by minimizing with a cost function in the hypercube instead of on the hypersphere gives better results.