By machine learning I mean the recognition and the generation ability. Like in the below photo. The Machine learning algorithm I imagine could recognize that it is a house and generate a plausible non pixelized image.
Generally for data streams. Perhaps a two layered approach could work. Where the first layer is more probability dependent. If that fails another deterministic algorithm is used. This way you have something like Lossy Error Correcting Code with a quality variable.