Born in Raleigh, North Carolina, Kyle Connelly is a student at North Carolina State University, working toward a double major in applied math and physics. His interests include just about everything, but he is focusing on artificial intelligence, and intends to study cognitive science in graduate school. He is currently working on applying chaos theory to machine learning systems, specifically reinforcement learning. He loves finding new perspectives on problems and concepts in math and science, and is excited to see what new perspectives the Wolfram Summer School has to offer.
Project: Describing Time Series Data
The goal of this project is to come up with a natural language description of time series data. Given time series data, this algorithm will extract relevant semantic information and generate a high-level description of the data in natural language. I first reduced the amount of data to analyze by fitting a sum of piecewise polynomials with local support to the raw data. I then came up with a “fuzzy classification” of the heights of the peaks and used this description to segment the fit into features, such as flat sections, fast growth, etc. Finally, I used some simple pattern matching and replacement rules to turn this “semantic information” into words describing the graph.
Main Results in Detail
I successfully set up a system to generate semantic information describing the general structure of time series data. This method can accurately describe local features and global trends, as well as the relative significance of each. My algorithm can describe a dataset of 2,000 points in around 500 ms. I have also implemented a simple system for converting this semantic information into descriptive words.
The system for going from semantic information to words is very simplistic and could be greatly improved. Additionally, a system for iteratively more detailed fitting could allow for a more detailed description without overfitting.