Classified Noise

Machine learning requires the right data to train the models. But sometimes this noise is not publicly available, as it may be a military secret.

“We work closely with partners that have a whole bunch of training data because the biggest issue we’ve got is having the appropriate training data,” said Philipp Stratmann, CEO of Ocean Power Technologies.

The firewalls between the different realms of classification pose a special challenge to those working to identify underwater sounds coming from databases that do not share security levels.

“The problem is the minute you start listening and looking for things that are classified, they sit in different classified buckets,” explained Stratmann.

Nevertheless, algorithms can help solve this problem.

By combining neural network models and adversarial ones, data can be created. This is called synthetic data.

“Combining experimental and synthetic data samples in the training phase produced additional performance gains,” according to research by the Federal University of Rio de Janeiro.