Model uncertainty quantification with dynamic parameters

A method, comprising:

• determining a first model output by applying a data sample to a model architecture having a distribution of dynamic model parameters using a first set of dynamic model parameters specified by a first sample from a probability distribution transformed by an invertible model;
• determining a second model output by applying the data sample to the model architecture using a second set of dynamic model parameters specified by a second sample from the probability distribution transformed by the invertible model;
• determining a quantified uncertainty of the data sample based on a plurality of model outputs including the first model output and the second model output;
• and
• determining that the data sample is out-of-distribution with respect to a training data set used to train the model architecture based on the quantified uncertainty.

A non-transitory computer-readable medium, the non-transitory computer-readable medium comprising instructions executable by a processor for:

• determining a first model output by applying a data sample to a model architecture having a distribution of dynamic model parameters using a first set of dynamic model parameters specified by a first sample from a probability distribution transformed by an invertible model;
• determining a second model output by applying the data sample to the model architecture using a second set of dynamic model parameters specified by a second sample from the probability distribution transformed by the invertible model;
• determining a quantified uncertainty of the data sample based on a plurality of model outputs including the first model output and the second model output;
• and
• determining that the data sample is out-of-distribution with respect to a training data set used to train the model architecture based on the quantified uncertainty.