MIT recently announced a 10-year partnership with IBM, centered around artificial intelligence and the Watson platform. Days before the announcement, STAT News published a report that discussed the challenges Watson faces in oncology, in actually helping physicians and patients. One of the big challenges is training and maintaining Watson’s knowledge base. Lack of data for training is also mentioned as a significant challenge in an MIT Technology Review article earlier this year about Watson.
Data science has become a huge buzzword, and definitely with all the huge amount of data available, many opportunities exist to gain new insights. But machine learning / artificial intelligence / deep learning (whatever you want to label it) all face the same challenges that they have always faced. Bias in training data leads to incorrect results, like with Google’s image tagging blunder in 2015. Lack of high-quality, “big enough” data sets make training difficult, as seen with Watson for Oncology. These challenges stem from the basic mathematics behind neural networks and other machine learning techniques — which all rely heavily on the training data input into the algorithms. To allow computers to learn like humans and be successful in different contexts, IBM’s Deep Mind is trying to make computers more like human brains. More advanced research similar to this, would be amazing to see out of the MIT / IBM collaboration, and I can’t wait to see what the researchers develop.