You could argue that training a model is essentially a lossy compression algorithm resulting a compressed artifact that can be non-deterministically decompressed.