Ever since computers were invented, we have wondered whether they might be made to learn. If we could understand how to program them to learn – to improve automatically with experience – the impact would be dramatic (cuote from the book: Machine learning. Tom Mitchell).
Definition of Machine Learning (book: Machine learning. Tom Mitchell):
A computer program is said to learn from experience E with respect to some class of tasks T and performance measure P, if its performance at tasks in T, as measured by P, improves with experience E.
Some insights of the video: Lecture 1. Machine learning. Professor: Tom Mitchell, that answer the question: Might computers be made to learn?
Machine learning. It’s about the question of how do we build computer programs that automatically improve what they’re doing through some kind of experience?
Machine learning is an interesting field, partly because there are a lot of interesting practical applications and algorithms, but partly because we also have the beginning of a theory to talk about properties of learning systems in general.
One of the most, I think, interesting sets of theoretical questions in computer science are really about questions about learning. So if you think about it, one of the most interesting questions that I can ask, in my opinion, is if I have some learning system, maybe it’s a computer, maybe it’s your brain, is there anything fundamental that relates the success of the learning strategy to properties of the training data that I present?
Example of never-ending learning:
NELL: Never-Ending Language Learning
So to have a well-defined problem in machine learning, we need to have three things specified. We need to have some performance task that the computer is performing, and some metric for what it means to do that better or worse. And then some kind of experience from which the program can train. So for example, you’re very familiar with spam filters, and you’re probably aware that spam filters are trained using machine learning.
But there’s some performance metric that, for example, you know what your own notion of spam is. And so there’s some, if you were, for example, to label a set of emails as ones that you consider to be spam or not, then we could say our performance metric we might use to make well-defined learning problem is to match as closely as possible the labels that you’ve assigned to the emails. That is for the program to learn to match as closely as possible your own labels.
We could say, for example, we could change the performance metric. Maybe you don’t really want to minimize the number of errors the spam filter makes compared to your gold standard. Maybe you actually are more worried about losing a good email into your spam folder than you are about accidentally looking at what is actually spam, so you might actually have a slightly different performance metric that you’d like to optimize that gives a bigger penalty if the program accidentally labels non-spam as spam than if it does it the other way around, so that would give you a different, again, well-defined problem.
We also could change the experience. So for example, so far we’re talking about training a spam filter by giving examples of emails with labels, spam not. Some computer programs, and we’ll learn about them this semester, can also take unlabeled examples, unlabeled emails, as additional data, and there are algorithms that can perform better, learning algorithms that can learn better if they’re given both those labeled examples and a whole bunch of unlabeled email, which seems kind of like magic.
It’s about how do we build computer programs that perform some task and improve their performance in some well-defined metric according to some type of experience.
Video: lecture 1 Machine Learning. Professor Tom Mitchell. Machine Learning Department. Carnegie Mellon University
© picture: Tom Mitchell. Machine Learning Department. Carnegie Mellon University.
Save the date:
ICML 2025
Forty-Second International Conference on Machine Learning
@Yolanda Muriel 
yolandamuriel.com 