Class-wise rationalization: teaching AI to weigh pros and cons
Natural Language Processing
What separates humans from the rest of the life on our planet? There are many factors, of course, but high on the list is the ability to form and convey complex ideas with a discernible language.
So if the goal is to maximize the utility of AI systems for humanity, they need to understand our natural mode of thought – and to communicate the way we do.
AI systems powered by neural networks have made great progress
in interpreting and mimicking language. But they’re still a long way from truly understanding language. We’re building AI systems that will cross the bridge from mimicry to comprehension. They’ll actually understand words, parse the meaning of rich ideas, and convert them into actual knowledge.
This new class of natural language processing systems will be powered by new types of neuro-symbolic systems that can understand both
the syntax and semantics of vast streams of language. They’ll connect complex language structures to the ideas they represent – and transcend today’s purely statistical approaches to language.
Neural-Symbolic VQA: Disentangling Reasoning from Vision and Language Understanding
Teaching language models grammar really does make them smarter
05/29/2019 - MIT News