Thus Spake Igoethustra
The biggest advance I made this week was recognizing that I’ve been going about things precisely backwards. Since the beginning of class I’ve focused primarily on finding the right algorithm(s) to study. I’ve been researching various supervised learning algorithms, but have been frustrated because 1) many of them are able to accomplish the same tasks, and 2) the differences come in the details of performance and circumstances of application. Because I currently have only a vague idea about the particular learning tasks I want an algorithm to accomplish, I’ve (naturally) had difficulty finding a suitable algorithm. As I’ve reread chapters from various texts, and synthesized information from conversations with Heather and the class however, I’ve come to appreciate that choosing an algorithm is but one part of designing a learning system, and is in fact the last step in that process (not the first). Since recognizing this, I’ve been focusing on the learning problems I would like solved, as well as on the potential tools that could help generating new questions which would be the target of future algorithmic investigation. This has led to a project timeline that focuses on machine learning algorithms only near the end of the semester. I’m very interested in getting feedback from Heather and from fellow students about this process though; I’m not sure it’s the way to go. Before jumping into the questions and timeline though, I’d like to (for the sake of my own clarity) review the process for designing a learning system as outlined by Mitchell.
The goal of any learning system is to improve, as judged by some performance measure P, at some learning task T, through experience (exposure to data) E. In order for any system to accomplish this goal, it must first be dealing with a learning problem that is itself well-defined, ie. T, E, and P are precisely spelled out, and there must a target function that stands for the knowledge that will be learned, a representation of than function (or model), and finally an algorithm that approximates that model. (to be continued….)
Potential Learning Problems
- Sleep Quality Prediction (multi-variate regression)
Task: predict sleep quality given ambient light, physical exercise, stress, and temperature.
Performance Measure: difference between predicted and actual sleep quality
Experience: Zeo ZQ score given fitbit, ambient light, temperature, nutrition, stress data.
Target Function Representation: q(s) = w0 + w1x1 + w2x2 + w3x3 + w4x4
Function Approximation Algorithm: stochastic gradient descent
- Cognition prediction
Task: predict mental acuity given continuous and discrete variables
Performance Measure: difference between predicted qualitative and measured acuity and actual qualitative and measured acuity
Experience:
Web-based stats package:
- descriptive statistics (max, min, range, mean, median, mode, sd, variance)
- historical data (manipulable date range, variables)
- scatter plot matrix of variables
Mobile training application
- queries about goals at fixed times during the day, logs answer, time, and location
- sleep quality (sleep well y n, sleep score from 1 to 5, times felt tiredd
Week 1 – Idea
Week 2 – literature review
Week 3 – build visualization/analysis program/literature review/gather data
Week 4 – build visualization/analysis program/literature review/gather data
Week 5 – build visualization/analysis program/literature review/gather data
Week 6 – build visualization/analysis/literature review/gather data
Week 7 – build visualization/analysis/implement first algorithms
Week 8 – build training application/implement first algorithms
Week 9 – implement second algorithm
Week 10 – implement second algorithms
Week 11 – overflow time
Week 12 – overflow time
If you enjoyed this article, please consider sharing it!
