Universal pre-training by iterated random computation

Peter Bloem

We investigate the use of randomly generated data for the sake of pre-training a model. We justify this approach theoretically from the perspective of algorithmic complexity, build…