Text this: Binarized neural networks converge toward algorithmic simplicity: empirical support for the learning-as-compression hypothesis