Sparse tree-based initialization for neural networks