Detecting egregious responses in neural sequence-to-sequence models

Tianxing He, James Glass

In this work, we attempt to answer a critical question: whether there exists some input sequence that will cause a well-trained discrete-space neural network sequence-to-sequence (…