Deep learning is a very popular topic today, and rightfully so, as it expanded the horizons of applicability of machine learning. But "shallow" models are often more interpretable and can lead to deeper insights and faster code. In general they are preferable over black box solutions for the problems that they can solve.

This talk aims to introduce the topic of machine learning as a tool for "everyday" programming, as a methodology of data-oriented problem solving. The talk will cover techniques like multidimensional data visualization, dimensionality reduction, nonlinear regression and symbolic regression. The presentation will show some examples from concrete rendering problems, but this talk will not be aimed at rendering engineers. Lastly, the talk will show that this methodology of iterative data exploration and modeling is useful to gain insights on the problem domain that often can lead even to solutions that do not require a learned model.

Takeaway

Attendees will learn why and how to apply data science reasoning and machine learning techniques to everyday programming problems. They will be able to better decide when a deep model is needed and when simpler techniques might prove to be superior. Also, they will start using data more to understand the complex behavior of today's problems and programs, instead of resorting to simplifying assumptions.

Intended Audience

This session is open to anyone who deals with programming problems. A bit of curiosity and knowledge of the basics of machine learning is a plus, but the speaker won't go into any deep mathematical theory or assume prior expertise with machine learning.