Training Machines to Answer Questions without Massive Data
An example of core AI research is in Ken Forbus’s Qualitative Reasoning Group, which is working toward training machines to answer questions without massive data.
Today’s deep learning systems can produce impressive results, but they require massive amounts of training data, can be brittle, and their results are not inspectable, making them hard to trust. People, by contrast, only require small amounts of data to learn, and can often explain their answers. The Qualitative Reasoning Group has developed Analogical Q/A Training as a more human-like way to train AI systems to answer questions. Based on a model of human analogical reasoning and learning, developed in collaboration with Dedre Gentner of NU Psychology, Analogical Q/A Training adapts a high-precision language understanding system to new tasks with only small amounts of data. Analogical Q/A Training has been shown to be competitive on several machine learning benchmarks, while requiring much less training data than deep learning systems. It has also been used in a deployed kiosk, which uses speech and vision to interact with visitors to answer their questions about Northwestern’s Computer Science Department. The group is continuing to explore these ideas, which could lead to making more trustable and trainable AI systems broadly available.
Learn more about Northwestern's Core AI Research
Back to top