That’s called generative recognition, and it’s the best way to do it (probably can be formally proved under some noise model assumptions) if you have a good model of the data and enough computational resources.