Natural language is valuable, but it is complex. With a 1,000 word vocabulary, a 15-word sentence can easily express more than 1e30 (a 1 with 30 zeros) different ideas. Today’s natural language processing is trained to bucket a sentence into one of a few thousand categories–which also means it only has a few thousand categories of things it can say back to you. When we simply classify text, we do not really understand it. Getting from 3,000 ideas to 1e30 ideas requires a breakthrough. This is what limits our ability to do data science on text.
With more than $33 million in funding, including one of the largest machine learning R&D contracts from DARPA over the past four years, Gamalon has developed a platform where models are Turing complete, compositional, transferrable; variables carry uncertainty; back-prop is just the tip of the iceberg; and people meaningfully interact with hidden layers in a model. This video demonstrates the breakthroughs that are coming in the field of natural language processing.
