Predictive modeling for dependency configuration in knowledge-augmented neural networks

A computer-implemented method for improving performance of a neural network, the method comprising:

• training, by a computer system, via machine learning, the neural network using a training data set to optimize a global objective for the neural network;
• identifying, by the computer system, a set of candidate knowledge sharing links for the neural network, wherein each candidate knowledge sharing link comprises a dependency from a digital knowledge source to a knowledge receiving node in the neural network;
• training, by the computer system, via machine learning, a plurality of network variants of the neural network, each version comprising a different configuration of the candidate knowledge sharing links;
• after the training of each network variant, determining, by the computer system, a measure indicative of performance for each network variant;
• fitting, by the computer system, a predictive model that estimates a relationship between one or more characteristics of the candidate knowledge sharing links and corresponding measures indicative of performance for the network variants;
• selecting, by the computer system, based on outputs of the predictive model, a configuration of the candidate knowledge sharing links to retain in the neural network;
• modifying, by the computer system, the neural network, to thereby generate a modified neural network, such that the modified neural network comprises the configuration of the candidate knowledge sharing links;
• and
• training, by the computer system, via machine learning, the modified neural network using the training data set to improve performance with respect to the global objective.