Tractable

Deep learning has just surpassed human accuracy in large scale vision (ILSVRC). Computers can now be made to see as well as humans. We're taking it from the research lab to the real world. We've automated the inspection of plastic pipe welds: this is the first time deep learning is commercially deployed to perform a human expert's task with superior accuracy. We're working on insurance. Our goal is to match human accuracy on it in the next 6 months. We're discovering what's holding back AI in the real world. The problem is the current requirement for plentiful, correctly annotated data: industrial datasets are noisily labeled at best. We believe in a pragmatic solution: a human-algorithm interface, powered by convnets, dimension reduction and active learning, that enables any human expert to transfer his expertise to the machine, rapidly, once and for all. We're researching and building it.