Collecting observations for machine learning

A non-transitory memory device storing device-executable code for training a model comprising a neural network including a generative network, the generative network having a latent vector as an input vector and a feature vector as an output vector, each element of the feature vector representing a different feature of a plurality of features, wherein weights applied by a node in the generative network are modelled as a probabilistic distribution, the code when executed by a processing apparatus directing the processing apparatus to perform operations comprising:

• obtaining an observed data point comprising a group of feature values, wherein within each group of feature values, each feature value is a value of a corresponding feature in the feature vector;
• training the model based on the observed data point;
• from amongst a plurality of potential unobserved features to observe, searching for a target feature of the feature vector which maximizes a measure of expected reduction in uncertainty in a distribution of said weights of the generative network based on the observed data point, wherein a measure of uncertainty comprises a measure of conditional mutual information in the weights of the generative network and a predicted classification based on the observed data point and the potential unobserved features;
• outputting, to an automated process, a request to observe a target data point comprising a value of the target feature, the request including a signal requesting the automated process collect the target data point of the target feature;
• receiving the target data point in response to the request;
• and
• further training of the model based on the received target data point,
• wherein either i) each of the features represents a different symptom of a subject, physical feature of the subject or other fact about the subject in a medical application and the model is used to make a medical diagnosis, or ii) the features represent different pixels of an image in a case of image recognition.

A computer for training a model comprising a neural network including a generative network, the generative network having a latent vector as an input vector and a feature vector as an output vector, each element of the feature vector representing a different feature, wherein a weight applied by a node in the generative network is modelled as a probabilistic distribution, the computer comprising:

• a processor;
• a memory device storing computer-executable code that, in response to execution by the processor, cause the processor to:
• obtain an observed data point comprising a feature value, wherein the feature value is a value of a corresponding feature in the feature vector;
• train the model based on the observed data point;
• from amongst a plurality of potential unobserved features to observe, search for a target feature of the feature vector, the target feature maximizing a measure of expected reduction in uncertainty in a distribution of the weight of the generative network based on the observed data point, wherein a measure of uncertainty comprises a measure of conditional mutual information in the weights of the generative network and a predicted classification based on the observed data points and the potential unobserved features;
• output, to an automated process, a request to observe a target data point comprising a value of the target feature, the request including a signal requesting the automated process collect the target data point of the target feature;
• receive the target data point in response to the request;
• and
• further train the model based on the received target data point.