<ul><li>An important challenge in machine learning is to predict the initial conditions under which a given neural network will be trainable.</li><li>A method for predicting the trainable regime in parameter space for deep feedforward neural networks (DNNs) is presented.</li><li>The method involves reconstructing the input from subsequent activation layers via a cascade of single-layer auxiliary networks.</li><li>The method shows promise in reducing overall training time and generalizes to residual neural networks (ResNets) and convolutional neural networks (CNNs).</li></ul>

Opening the Black Box: predicting the trainability of deep neural networks with reconstruction entropy

Discover more