Artificial intelligence scientists have raised concerns about the information retained by neural networks, ZDNet reports. Google Brain Research Scientist Nicholas Carlini published a paper in which he and his colleagues discuss how neural networks may retain portions of a dataset used to train it to generate text. The paper poses malicious entities could infiltrate a network for sensitive data, such as credit card and Social Security numbers. "Ideally, even if the training data contained rare-but-sensitive information about some individual users, the neural network would not memorize this information and would never emit it as a sentence completion," the paper states. "Unfortunately, we show that training of neural networks can cause exactly this to occur unless great care is taken."
Full Story