CVPR Daily - 2018 - Thursday

biologically inspired. Think of the way our nervous system works, the way neurons communicate and so on, via spikes. Is there anything you can tell us about IBM that we might not already know? One thing that is probably very important is that we are making tremendous progress in the field of AI today. Most of this progress is really what we would call narrow AI. Narrow tasks. Where the success is coming from is generally when we have a large amount of training data. We think of narrow AI as we have a small number of large data problems and deep learning is doing well. Where IBM is particularly focused is the enterprise space and industry problems. The nature of the problems in the enterprise space is we have a very large number of small data problems. When we apply deep learning just out of the box, it doesn’t work the same way it does in narrow AI. It’s putting new emphasis on learning more from less data. We need techniques of transfer learning –greater transferability, better exploitation of knowledge and reasoning combined with learning. We see this as very important as we go forward. More than that, as we really want to apply techniques like deep learning for industry applications, the majority of them are decision support. There are people who need to make decisions. Think about a healthcare application – a doctor who is trying to use AI tools to make a better diagnosis of a patient or treatment recommendations. That model may be accurate, which is great, but if the doctor and patient cannot understand how the computer is reaching whatever decision, then this may be a big gap to really rely on its output. It’s putting emphasis on explainability. We know that there are also issues about bias in how we train, and often human bias finds its way into bias during learning. Security of models is also very important. How do we know the models aren’t poisoned somehow or corrupted? This whole space of problems we contrast to the narrow AI field and we think of this as more a broad AI. Although that is not necessarily the accepted term, but it’s saying that there’s still a lot of gaps we need to bridge here to really make AI 6 Thursday John R. Smith - IBM