We have spent the last twenty-five years developing technology for performing large-scale inference on Bayesian networks. Our motivation for this was to support the use of these networks for medical diagnosis – long a focus of the Artificial Intelligence community. We have developed and tested detailed models of infectious diseases and believe the system to be ultimately capable of diagnosing all human disease. We have successfully evaluated networks containing more than 30,000 nodes and networks with nodes conditioned by 5,000 other nodes.
The current Covid-19 pandemic has prompted us to announce this technology now in the hope that a system can be developed and deployed to aid globally in computer-assisted medical diagnosis and the fight against Covid-19. Such a system would initially address primary care which would also include all quarantinable diseases that could result in a pandemic. The technology could not only diagnose diseases like Covid-19, but also detect the emergence of new, as-yet unknown, diseases and thus possibly prevent future pandemics that might be caused by them.
We are offering a new service whereby Bayesian Networks (BN) are used to produce synthetic training data for Deep Learning with Artificial Neural Networks (ANN). Both network types can take a number of numeric input values and perform a numeric transform on them to produce a number of numeric output values. For a BN, the transform is provided by Bayesian probability analysis. For a certain type of ANN, the transform is “learned” from an examination of sample transformations where the inputs and outputs are known. This type of learning is known as “supervised learning” and the resulting ANN is known as a “classifier”.
Executing a transform by a BN can be slow compared to a similar transform executed by a trained ANN. A BN requires no training – every possible transform is specified by the structure of the network. An ANN can require extensive training – sometimes millions of samples. The structure of a BN can be changed easily while retraining an ANN can be very slow.
The requirement for large amounts of data for training an ANN can be an insurmountable impediment for the use of an ANN for certain applications. For example, consider detecting an impending melt-down in a nuclear reactor. Fortunately, in real life, there are very few examples. Unfortunately, that means there are too few to train an ANN. However, the problem can be modeled using a BN incorporating only the probabilistic relationships among reactor components. Since the BN and the ANN would solve the same problem, the BN can be used to generate synthetic examples of the required training data. This has the added advantage of training on inputs that have never occurred but are possible.