Supervised, unsupervised and reinforcement learning
The three learning modes and what separates them, labelled examples, unlabelled data or a reward, the agent-environment loop, self-supervision and badly specified rewards.
- 25 cards
- 6 topics
- Free
The outline, topic by topic
Each topic is weighted the way the exam weighs it.
- Concepts
- Questions
- Sort and link
- The numbers
- Scenarios
- Understand
A few cards from the deck
Supervised learning: what is it?
Show the answer
The mode in which the model learns from examples that each carry the right answer, known in advance.
The right answer of each example is called its label; it comes from human annotators or from records, such as the price of a past sale. Labelling is costly, and the model will not exceed the quality of these labels.
An agent has a discount factor of 0.1. What behaviour should we expect from it?
- It favours immediate rewards, even smaller ones
- It aims for the biggest possible reward, however distant
- It explores at random, because it does not remember its past rewards
- It learns more slowly, because each update is divided by ten
Show the answer
It favours immediate rewards, even smaller ones
A reward received two steps later is multiplied by 0.1² = 0.01: it hardly counts. The factor sets the agent's horizon, not its learning speed.
At each step, the … chooses an action; the … returns a new state and a reward.
Show the answer
At each step, the agent chooses an action; the environment returns a new state and a reward.
This loop of observing, acting and receiving a reward is the same for a robot, a video game or a go program.
Unsupervised learning: what is it?
Show the answer
The mode in which the model receives data with no expected answer and looks for a structure on its own: groups, main axes, anomalies.
Grouping customers into segments is the typical case. Nothing says whether the groups found mean anything: a human has to interpret them.