Quiz 00.2 on Machine Learning

Q1: What is Machine Learning?

Which could be considered examples of Machine Learning ?

An international team of oncologists has reviewed the empirical evidence on breast cancer and produced a decision support tool for practitioners that prescribes the best medical act to take on patients based on chemical, biological and histological characteristics of the tissues and the patient genetic, physical and medical condition

Your preferred photo editor tool function to cartoonise your photos

Your content provider suggests you the films it thinks you may like based on the films that you have already decided to watch and the review given by the other users that have viewing behaviours similar to yours

Your phone automatically composes a video by selecting the most beautiful photos

RESOLUTION

To be considered ML, an algorithm must learn the solution of the problem from the data. In the oncologist example, it is humans that define the outcomes, even if this decision is data-driven. Similarly, in the photo editor, the parameters of the filters that are applied (convoluted) to the photo to cartoonise it (or making whatever other effect) are not learned by the algorithm itself but explicitly engineered. We will see in the NN unit how this contrasts with Convolutional Neural Networks, where the weights of these filters are instead learned.

The correct answers are:

Your content provider suggests you the films it thinks you may like based on the films that you have already decided to watch and the review given by the other users that have viewing behaviours similar to yours
Your phone automatically composes a video by selecting the most beautiful photos

Q2: Kind of machine Learning tasks

Which of the following statements are correct ?

In unsupervised (machine) learning the objective is to learn the relation between some inputs and some outputs from a sequence of pairs (inputs, outputs) provided to the algorithm

In supervised machine learning tasks the objective is to learn the relation between the provided features (inputs) and the provided labels (outputs)

In reinforcement learning tasks, the algorithm must find the best actions for a certain agent to perform given the different states of the world, the rewards that the agent is given at each possible state and the probabilities to reach the various states given the departing state and the available actions

None of the (other) sentences is correct

RESOLUTION

All sentences are wrong. Unsupervised machine learning, by definition, includes algorithms for which we do not provide examples of a "correct" output for the different inputs. We try instead to find a pattern, a structure in the data itself.

Outputs are provided in supervised tasks. However the sentence reported is wrong, because we are not interested much in the relationship between the provided inputs and outputs, but in finding a generic relationship between the inputs and the outputs for the population from which the data arise.

Finally, also the sentence on reinforcement learning is wrong, because in reinforcement learning we don't know the rewards associated to each state nor the probabilities to reach the various states associated to each action. This is what the algorithm needs to discover (learn) by start "playing" autonomously.

The correct answer is:

None of the (other) sentences is correct