Übersicht
Möchte man einen Algorithmus entwickeln, gibt es in der Regel zwei große Ansätze, die je nach Situation und Endanwendung zum Einsatz kommen:
| Topic | Supervised | Unsupervised |
|---|---|---|
| 🎯 General Focus | I/O is labeled (we already know what we are searching for) | I/O is unknown. AI has to decide on the variables. |
| 👨🏻🎓 Learning Style | Big Set of Reference Data - More Precise, but more work for the Human | Tries to detect patterns from the Input. The Human must give “sense” to the Output (+ adjusting Parameters if needed) |
| 🏁 Goals | Target Results are already known. The AI helps in sorting the Data-Set and predicts the outcome | Out of a big set of Data and Variables, the AI tries to find correlations. |
| 🤖 Application | SpamScanner, Forecasts, Pricing | Finding Anomalies, Recommendation Systems, Customer Personas, Medical Engineering (Images) |
| 🕸 Complexity | Rather “simple” Machine Learning (R / Python) | High Complexity: needs a lot of Rechenpower and Data to be trained accordingly |
| ⚠ Drawbacks | Needs a lot of time to develop. Human experts need to evaluate needed I/O first | Creates Random Results and need Human experts to give “sense” to the I/O the AI chose. |
Sonstige Vermerke
- Es gibt auch einen Mittelweg, der als Semi-Supervised Learning bezeichnet wird (https://en.wikipedia.org/wiki/Semi-supervised_learning))
- Wie groß muss Trainings- und Testsample sein, um einen Algorithmus entsprechend zu trainieren?
Weiterführende Literatur
- Schuler, Nikolai Statistik & Mathematik für Data Science & Data Analytics (https://www.udemy.com/course/statistik-data-science/))
zettelkasten Links
- ZK-006-PSY-Intelligenz - hoch vs. niedrig
- ZK-045-PSY-Alpha bis Omega - Persönlichkeiten to lead or follow suite
- ZK-072-EDU-Lernen mit vs. ohne Regelwerk (bsp. Musik und Schach)