Associating a population descriptor with a trained model
A model may be trained on a training dataset, e.g., for medical image processing or medical signal processing tasks. Systems and computer-implemented methods are provided for associating a population descriptor with the trained model and using the population descriptor to determine whether records to which the model is to be applied, conform to the population descriptor. The population descriptor characterizes a distribution of the one or more characteristic features over the training dataset, with the characteristic features characterizing the training record and/or a model output provided when the trained model is applied to the training record. For instance, the model may be applied only to records conforming to population descriptor, or model outputs of applying the model to non-conforming records may be flagged as possibly untrustworthy.
1. A system for processing training records of a training dataset, comprising:
a data interface for accessing:
a training dataset for a trained model, the training dataset comprising a plurality of training records, wherein the trained model is trained on the training dataset to provide a model output when applied to a record;
a processor subsystem configured to:
determine one or more characteristic features for a training record of the training dataset, said one or more characteristic features characterizing the training record, said characteristic features being based on the training record and/or metadata associated with the training record;
determine a population descriptor characterizing a distribution of the one or more characteristic features over the training dataset, wherein said determining does not access intermediate outputs of the trained model;
associate the population descriptor with the trained model to enable an entity which applies the trained model to at least one record to determine whether the at least one record conforms to the population descriptor by determining the one or more characteristic features for the at least one record and by determining a deviation of said one or more characteristic features from the distribution of said characteristic features over the training dataset.
2. A system as in claim 1 , wherein the population descriptor comprises probability data indicative of a probability distribution of the one or more characteristic features over the training dataset.
3. A system as in claim 1 , wherein the population descriptor comprises a one-class classifier trained on the one or more of characteristic features of the training records to enable the entity applying the trained model to classify the at least one record as being conformant.
4. A system as in claim 1 , wherein a training record comprises an image, the processor subsystem being further configured to determine the one or more characteristic features of the training record by extracting a characteristic feature from the image and/or by determining a characteristic feature from metadata associated with the image.
5. A system as in claim 1 , wherein a training record comprises an image and the trained model is trained to provide a segmentation of the image, the processor subsystem being further configured to determine a characteristic feature of the training record by extracting a shape descriptor of the segmentation of the image.
6. A system as in claim 1 , wherein the characteristic features for the training record are features of the training record that were used as input when training the trained model, or metadata that was not used as input when training the trained model.
7. A system for processing at least one record to which a trained model is to be applied, comprising:
a data interface for accessing:
a population descriptor associated with the trained model, the trained model having been trained on a training dataset, the training dataset comprising a plurality of training records, the population descriptor characterizing a distribution of one or more characteristic features for the training records over the training dataset, the one or more characteristic features of a training record characterizing the training record, the one or more characteristic features being determined based on the at least one record and/or metadata associated with the at least one record, the population descriptor having been determined based on the characteristic features without accessing intermediate outputs of the trained model;
the at least one record to which the trained model is to be applied;
a processor subsystem configured to:
determine the one or more characteristic features of the at least one record;
determine a deviation of the one or more characteristic features of the at least one record from the distribution of said characteristic features over the training dataset to determine whether the at least one record conforms to the population descriptor;
when the at least one record does not conform to the population descriptor of the trained model, generate an output signal indicative of said non-conformance.
8. A system as in claim 7 , further comprising an output interface for outputting the output signal to a rendering device for rendering the output signal in a sensory perceptible manner to a user.
9. A system as in claim 8 , wherein the output interface is further configured to output a deviation of one or more determined characteristic features that resulted in the non-conformance to the population descriptor.
10. A system as in claim 7 , wherein the data interface is configured to access population descriptors associated with multiple trained models, the processor subsystem being configured to:
determine whether the at least one record conforms to a population descriptor associated with a respective trained model; and
in case of conformance, select the respective trained model of the multiple trained models for being applied to the at least one record.
11. A system as in claim 7 , wherein the processor subsystem is further configured to adapt a model parameter of the trained model in response to the output signal.
12. A computer-implemented method of processing training records of a training dataset, comprising:
accessing:
a training dataset for a trained model, the training dataset comprising a plurality of training records, wherein the trained model is trained on the training dataset to provide a model output when applied to a record;
determining one or more characteristic features for a training record of the training dataset, said one or more characteristic features characterizing the training record, said characteristic features being determined based on the training record and/or metadata associated with the training record;
determining a population descriptor characterizing a distribution of the one or more characteristic features over the training dataset, wherein said determining does not access intermediate outputs of the trained model;
associating the population descriptor with the trained model to enable an entity which applies the trained model to at least one record to determine whether the at least one record conforms to the population descriptor by determining the one or more characteristic features for the at least one record and by determining a deviation of said one or more characteristic features from the distribution of said characteristic features over the training dataset.
13. A computer-implemented method of processing at least one record to which a trained model is to be applied, comprising:
accessing:
a population descriptor associated with the trained model, the trained model having been trained on a training dataset, the training dataset comprising a plurality of training records, the population descriptor characterizing a distribution of one or more characteristic features for the training records over the training dataset, the one or more characteristic features of a training record characterizing the training record, the one or more characteristic features being determined based on the at least one record and/or metadata associated with the at least one record, the population descriptor having been determined based on the characteristic features without accessing intermediate outputs of the trained model;
the at least one record to which the trained model is to be applied;
determining the one or more characteristic features of the at least one record;
determining whether the at least one record conforms to the population descriptor by determining a likelihood of the one or more characteristic features of the at least one record occurring according to their distribution over the training dataset;
when the at least one record does not conform to the population descriptor of the trained model, generating an output signal indicative of said non-conformance.
14. A non-transitory computer-readable medium comprising transitory or non-transitory data representing instructions arranged to cause a processor system to perform the computer-implemented method according to claim 12 .
15. A non-transitory computer-readable medium comprising transitory or non-transitory data representing a trained model and a population descriptor associated with the trained model, the trained model having been trained on a training dataset, the training dataset comprising a plurality of training records, the population descriptor characterizing a distribution of one or more characteristic features for the training records over the training dataset, the one or more characteristic features of a training record characterizing the training record, the one or more characteristic features being determined based on the at least one record and/or metadata associated with the at least one record, the population descriptor having been determined based on the characteristic features without accessing intermediate outputs of the trained model.
16. The computer-implemented method of claim 13 , wherein the population descriptor comprises probability data indicative of a probability distribution of the one or more characteristic features over the training dataset.
17. The computer-implemented method of claim 13 , wherein the population descriptor comprises a one-class classifier trained on the one or more of characteristic features of the training records to enable the entity applying the trained model to classify the at least one record as being conformant.
18. The computer-implemented method of claim 13 , wherein a training record comprises an image, further comprising determining the one or more characteristic features of the training record by extracting a characteristic feature from the image and/or by determining a characteristic feature from metadata associated with the image.
19. The computer-implemented method of claim 13 , wherein a training record comprises an image and the trained model is trained to provide a segmentation of the image, further comprising determining a characteristic feature of the training record by extracting a shape descriptor of the segmentation of the image.
20. The computer-implemented method of claim 13 , wherein the characteristic features for the training record are features of the training record that were used as input when training the trained model, or metadata that was not used as input when training the trained model.