Are Neural Networks Cheating?
Neural networks produce state-of-the-art performance in several tasks these days. But how are they doing it? Or are they actually cheating their way through? In this talk I will present research that is being done at the Computational Language Understanding Lab of the Computer Science department of University of Arizona. Here, I will first introduce how neural networks work, from a math perspective, and then we will look deep into the workings of neural networks using natural language processing as a tool. I will then present how neural networks, many-a-times, rely on several subtle statistical patterns in the data to produce these state-of-the-art results. Further, I will also present a solution we have developed in our lab to counter this 'memorizing' habit of neural networks. The proposed solution, called Confluence Learning, uses data distillation and model distillation approaches over fact verification datasets. By using different delexicalized views of the data and encouraging the models to learn from each other through pairwise consistency losses, we prevent the neural networks from relying on such dataset artifacts, but instead learn the true underlying semantics of a given dataset.
Details about the presenter: Mitch is a final year PhD candidate at University of Arizona’s Department of Computer Science, where he is advised by Professor Mihai Surdeanu. He is broadly interested in the implementation of neural networks for natural language processing. His current research focuses on fact verification with an emphasis on domain transfer and has published in many prestigious conferences in the field including EMNLP. Mitch earned his bachelor's degree in Physics and Engineering, along with a Masters degree in Physics, from Birla Institute of Technology and Science(BITS), Pilani, India. He did his second masters in Computer Science at the University of Arizona. His other interests include network security and penetration testing.