Lecture 16: Three Learning Principles
16-1 Occam's Razor
Simple hypothesis H (few parameters), simple model (small number of hypotheses)
16-2 Sampling Bias
In the 1948 presidential election, telephone sampling results did not match actual results. Reason: Telephones are not widespread and the wealthy are being sampled.
When the sampling is biased, it will affect the learning results.
For example: the recommendation system is based on the time axis and cannot be randomly sampled; the system of bank credit cards only has information on whether the issued credit cards are paid back on time, and there is no record of unissued cases. Sampling learning cannot determine whether a credit card should be issued.
16-3 Peeking at the data
The data is polluted by its own selection decisions.
16-4 Summary
so! ! ! Finished watching the video of Machine Learning Cornerstone! Next, I should not watch machine learning techniques for the time being, and I am going to watch a total of 16 weeks of videos in the "deep learning" series on coursera. Don't be in a hurry to finish it~ It's enough to solve it on time on the weekday evening this semester, and wait for the financial aid of coursera to be approved.