Our machine learning solution stems from roof drainage systems. It warrants that large buildings like stadiums or shopping malls do not collapse during heavy rainfalls. Choosing the “right” diameters is difficult, requiring intuition and hydraulic expertise.
We use neural nets to improve the existing heuristic planning tools, reducing the fail rate from 24% to 6%. We show the pitfalls of generating data from a limited data set (“big data” is missing in the “small world” of roof drainage systems, more data is created only as new buildings are planned) as well as why “more” may not be “better”. We conclude by showing how changing the learning goal was instrumental to the solutions success.
Required audience experience: Basic familiarity with machine learning helpful, but not required (level 100).
Objective of the talk:
• Beware of good results
• Deep neural nets need domain knowledge too
• Data extraction is king / big is not enough
• Continuous improvement in production
• Tools are not the problem (anymore)
Keywords: Azure Machine Learning, neural networks, pitfalls of randomized data generation, data extraction, label engineering, operations of a ML based product
You can see the slides for Christian’s talk via the link below: