There are all sorts of things you could cover — machine learning is an extremely interesting and growing field, with many different approaches and tools you could explore.
Some of the more popular tasks for machine learning algorithms include classification, regression and clustering. Not every application fits into this grouping (e.g. AlphaGo, the Go playing program which beat Lee Se-dol).
I might suggest starting with Naive Bayes classifiers. Your motivating example here could be spam filtering; Naive Bayes classifiers were used in some of the first 'learning' spam filters, dating back to the 1990s. If your students are familiar with Bayes' theorem, the assumptions of NB classifiers shouldn't be a huge leap (you may find this derivation interesting).
There are also various other interesting options such as random forests (apparently used by Quora to find duplicate questions), support vector machines and so forth which could be explored potentially.
Many of the most 'fashionable' techniques involve neural networks, and I would recommend spending a decent amount of time with the theory. Unlike some of the classifiers and regressors I mentioned earlier, neural networks tend to be a bit more involved — a Naive Bayes classifier essentially just needs to be given some data in an appropriate form and is "plug and play", so to speak.
The (free) online book Neural Networks and Deep Learning gives a reasonably accessible explanation of neural networks in their various forms (starting with perceptrons, leading towards sigmoid neurons, gradient descent, backpropagation, deep learning, etc.). As you can imagine, there is an immense amount of content even in just the field of neural networks.
Likely you won't want to cover all of the book's topics, but you don't need to go too far to reach another interesting motivating example: classifying the MNIST dataset of handwritten digits. That's something which is relatively hard traditionally, but much simpler with the usage of a neural network.
Note that I've discussed a lot without ever actually explaining how you could write any programs. Of course, any machine learning class would be a little incomplete if your students had never actually applied their skills.
Since your students know Python, I can highly recommend scikit-learn. It's widely regarded as an exceptionally good ML library, and the API is very friendly — you can almost treat it as LEGO, and plug together the pieces needed to do your classification/regression/clustering... even without really knowing how each piece works entirely. You can get functional (though not exactly effective) solutions simply by connecting the appropriate pieces in a pipeline.
For example, if you wanted to classify some texts, you'd just:
- collect some data
- setup a
Pipeline of a CountVectorizer and a classifer such as LinearSVC
- train it, and then test it.
There's an awful lot of stuff included, but then again, all the batteries you'll ever need are included, too.
For playing with neural networks, Keras might be worthwhile.
That said, your students may derive some value from implementing the algorithms themselves, at first. The key is to be able to find truly motivating problems:
Struggling with a project you care about will teach you far more than working through any number of set problems. Emotional commitment is a key to achieving mastery. (source)
It's not always easy if you're supervising many students, but it is worthwhile advice to consider.
