5 things ML can teach us about learning

In the last few months, I have had the chance and the pleasure to teach some Machine Learning concepts in the internal Data Science Bootcamp at IE University. During this time I’ve been thinking that in an educational institution like ours, learning hashtag#machinelearning has an interesting flavour: discovering how some experts have gone about a familiar problem like hashtag#teaching, simply to machines instead of people.

So here you have 5 things that IMHO ML can teach or remind us about hashtag#learning:

• The most desired models don’t simply “remember” given data, rather generalize well on unseen data → Effective learning shouldn’t be about remembering everything we’re given (books, slides, etc.), rather grasping the underlying patterns/concepts.

• Nobody actually wants nor trusts a model with 100% accuracy → Making absolute no mistakes isn’t the goal of learning, we should leave a space for error

• Accuracy isn’t the only metric, nor exists an ultimate, all-inclusive metric → There isn’t only one way to evaluate learning and every score will miss something

• If you overfit, simplify your model; if you underfit add more data or more features → If you learned things by hearth and you struggle in the exam, you should get rid of details and focus on the big picture; if you struggled both at home and in the exam, get your hands on more books and information

• All models need to be retrained from time to time → Learning isn’t a one-off, we constantly need to update our knowledge and keep on learning