For those just getting started with advanced scientific computing techniques, here are four steps to efficiently turn data into decisions with business value.
Author: Ryan Swindeman, Scientific Software Developer
In the 4 minute video below, Enthought scientist Ryan Swindeman puts data into context as foundational to any digital transformation initiative, setting out four fundamental steps for data to science problems.
1. Data Preparation (or Data Conditioning): This is the essential, first step in a digital project. Data must be clean and accessible. Access to data must be quick, and reliable. The data must be cataloged or categorized, so that there is consistency in how it is reached and integrated into projects. Data preparation must be in service of addressing a business need or objective, to solve a specific problem, and not be a case of ‘we need to organize our data’.
2. Data Visualization: Visualizing data is important as a starting point to understanding a problem. This involves looking at the data in its native domain, identifying trends, and from there possibly transforming it to a different domain, cross-plotting to look for relationships, or running statistics as a way to discover features. Visualization is also a reliable way to increase efficiency in problem-solving. The understanding gained through visualization is essential for deep learning – if you do not understand the underlying trends or relationships in the data, you will not understand the outcomes produced by any AI/ML/Deep Learning.
3. Modeling and Optimization: This step uses the underlying dynamics or physics of the problem, and the applications are endless. (In geophysics, this is often called forward modeling and inversion.) Most critically, modeling and optimization allows scientists to prove (or disprove) hypotheses very quickly, enabling teams to test, iterate and change strategy, often resulting in problems being solved quickly.
4. AI/ML/Deep Learning: These advanced computing techniques are related, and differ in important ways. Unlike modeling and optimization, or inversion (which is a physics-based approach), AI/ML/Deep Learning is a data-driven approach. These techniques are beneficial if forward modeling and optimization are not possible because of a lack of understanding of the underlying physics, or if the physics leads to too many approximations. The problem-solving and analytical power of AI/ML/Deep Learning becomes obvious in pattern recognition or texture analysis.
These four steps provide a robust sequence for solving problems using data, whether a small set or large, fundamental to digital transformation projects.
About the Author
Ryan Swindeman, Scientific Software Developer at Enthought, holds a M.S. in geophysics from the University of Texas at Austin and a B.S. in physics from the University of Illinois at Urbana-Champaign, with graduate research in computational seismology.