The awakening of artificial intelligence
Predictive Analytics World is the conference in German-speaking countries when it comes to data science and predictive analytics. This year's theme "The awakening of artificial intelligence" arouses high expectations; if "artificial intelligence" is to be understood not just as a marketing word for "data science", then only a focal topic in the area of deep neural networks can be implied. However, it is not yet a focal topic, as revealed in the program - of the 18 lectures on the first day, three, if at all, focused on deep neural networks.
Deep Recurrent Networls - the program highlight on the first day
One of the lectures stood out, however, namely the one by Dr. Ralph Grothmann: "Deep recurrent networks: Theory and industrial applications at Siemens". The level of interest was high. All of the 35 or so seats in the room were taken, about twenty people listened to the presentation while standing, and others were turned away at the door because of overcrowding. The lecture's conventional introduction involving amply drawn out distinctions between "predictive vs. prescriptive analytics" was soon followed by much more exciting content. Dr. Grothmann presented very sophisticated applications especially for large-scale industrial facilities, but also for predicting door failure in Inter-City Express trains, for example. The applications were sooner predestined for classical data-science methodology, though it is somewhat doubtful whether some could still have been solved with reasonable effort by means of "classical methods" (ensemble models, support vector machines, statistical regression approaches, time series analysis, etc.).
More interesting than the applications, however, was the methodology with which they were handled. Unlike many protagonists in the deep-learning environment who often proceed according to the motto "why should I care about business logic, if you give me enough data", Dr. Grothmann stressed the importance of understanding the task's logic, mostly based on engineering and physics. His approach follows the idea of constructing the neural network in close orientation with the application logic. Typically, the resulting networks (depending on the number of input variables) are significantly smaller than the often huge networks used, for example, in image processing, and tend to involve hundreds to thousands of nodes, not the hundreds of thousands found in more typical deep-learning applications.
The lecture was completed by a very interesting question-answer session which provided, among other things, also very pointed insights into the avoidance of overfitting ("regularization is the old school answer!").
Conclusion - Day 1
All in all, this presentation alone was worth the journey to Berlin. In my opinion, the approach of not favoring mountains of training data by ignoring a knowledge of the application's problem, but instead including it in the construction of the neural networks to be employed, is very useful and promising if deep learning is to be introduced into classical predictive analytics, and I am looking forward to testing it in my own projects.