Commentary on "Reliable Reasoning"

This book is a wonderful redux to a time in the early half of the 20th century when statistical learning theory was just developing, and when new methods and concepts were being discovered. By the end of the 1970s learning theory had been sidetracked by the need for practical results that were consistent with the prevailing dogma of AI (artificial intelligence). Proponents of AI had posited that human thought was propositional in fact, specifically related to some unknown logical form. Consequently, much of the technology and computer science was devoted to developing the LISP or Prolog language that would somehow be more consistent the “programming language” of the mind. Early work in Machine Learning (1970s-1980s) focused on algorithms that had some functional relationship to some fundamental human task: categorization, perceptual object recognition, language understanding, reasoning, for example. But the algorithms were often too narrowly focused on these types of tasks and were termed as “brittle”, in that small variations in the input conditions caused the algorithm to completely fail. One of the often repeated stories that was associated with the abrupt decline of AI during the late 1980s, was a DARPA project involving autonomous navigation of a tank in off-road environments. With the assembled VIPs of DARPA project Managers, Directors and military brass, the tank performed flawlessly through a set of standard obstacles and off-road variations, until, the sun went behind the clouds, causing the tank to immediately take a right turn into a tree continually ram the tree over and over again. After this event and others of a similar vein, much of the AI funding at major centers (MIT, CMU) was almost entirely cut. Insult was added to injury when a graduate student at CMU (Dean Pomerleau) developed a modest neural network (he dubbed ALVNN-autonomous learning vehicle Neural Network) that could be trained under various weather, road and obstacle conditions by driving the vehicle for a few hours in diverse conditions and thereafter would perform flawlessly on the same tasks as the AI programmed tank including varied lighting conditions which proved