Deep learning has brought immense improvements, but it is also very opaque. Breakthrough papers present new architectures that performed better on benchmark data sets, but there is little scientific understanding why that is so.