- Roy Schulte
14 November 2016.
The opinions in this document are my own, and do not represent the position of my employer or any other company.
AI emerged as a major theme at the recent Strata+Hadoop conference in New Yorkin September, 2016, along with event stream processing and more-traditional forms of data and analytics. Gary Marcus, CEO of Geometric Intelligence and professor of psychology & neural science at NYU, gave a terrific presentation about AI on 29 September.
He pointed out that AI is still effective only in narrow domains, and it makes a lot of mistakes. Roomba, the intelligent vacuum cleaner, is not nearly as smart as Rosie the Robot from the Jetson cartoons in the 1960s. The MIT AI project Eliza back in 1965 could already fool people into thinking that it was a real therapist; Siri is smarter but is still limited. A current, well-trained deep learning system thought that a traffic sign with some stickers on it was a refrigerator full of food. A self driving car in New Jersey needed two human interventions for a daily commuting trip – which is an unacceptable level of error.
AI has three main limitations:
- Poor performance in the “long tail” rare cases where you don’t have a lot of data on which to train.
- Difficult to debug, revise and verify.
- Can’t represent causal knowledge and apply it, and can’t integrate abstract knowledge. For example, a physical model of how something should work is not part of the learning process for a deep neural network (DNN).
Gary’s solution (summarized simplistically – he actually said a lot more) is to expand the number of algorithms used for each AI application. The human brain has maybe a thousand kinds of neurons, but DNNs have only one. Human neurons differ structurally, functionally, genetically, and in how they form connections.
Gary was not the only speaker on AI. Among others, Martin Hall from Intel also said we are entering the age of AI. He mentioned neural networks, Internet of Things applications, text and video analytics. Progress in AI is about openness and collaboration between software components and between software and hardware.