Training Dynamics and Neural Network Performance

Fingerprint image quality

“…The theory behind the machine learning techniques used in this program is discussed in [17]. It trains a 3-layer feed forward nonlinear perceptron model.…”

“…This is common in classification problems where large samples of rare classes may not be available. We have compensated for lower samples of poor quality in our training set by setting "class-weight" parameter of our implementation of neural network [17]. Our implementation of neural network accepts class prior probabilities as a training parameter.…”

Effect of resolution and image quality on combined optical and neural network fingerprint matching

“…The matching networks discussed m this section were trained using a dynamically modified scaled conjugate gradient method presented m [2]. In [2], we demonstrated that performance equal to or better than Probabilistic Neural Networks (PNN) [19] can be achieved with a single three-layer Multi-Layer Perceptron (MLP) by making fundamental changes in the network optimization strategy.…”

“…We study matching of the inked fingerprints using global optical correlations [1] and partial optical correlation features and a system of neural classification and matching networks [2]. Images of three difierent resolutions are tested to determine the effect of image resolution and quality on matching accuracy.…”

“…These two networks function in a way similar to the binary decision networks discussed in [7]. For this particular problem, the network training is strongly dependent on regularization and pruning for accurate generalization [2].…”