Searching for minimal optimal neural networks

Large neural network models have high predictive power but may suffer from overfitting if the training set is not large enough. Therefore, it desirable to select an appropriate size for networks. The destructive approach, which starts with a architecture and then reduces using Lasso-type penalty, has been used extensively this task. Despite its popularity, there no theoretical guarantee technique. Based on notion of minimal networks, we posit rigorous mathematical framework studying asymptotic theory We prove that Adaptive group Lasso consistent can reconstruct correct number hidden nodes one-hidden-layer feedforward networks probability. To best our knowledge, first result establishing