Types of learning
Alongside supervised and unsupervised, this is the third main type of machine learning, where an agent learns from rewards.
What is reinforcement learning?
This is the kind of supervised task used to predict a continuous number, such as a house price.
What is regression?
In a classification tree, this is the name for the very first split at the top, before any branches.
What is the root node?
This unsupervised technique groups similar data points together, like Spotify grouping songs by usage patterns.
What is clustering?
Described as "white-box," this kind of model is directly understandable on its own, like a linear regression or a decision tree.
What is an interpretable model?
Supervised learning requires these correct answers to be attached to each training example.
What are labels?
This basic regression model, often called "Ordinary Least Squares," minimizes the sum of squared residuals
What is linear regression?
This overall classification metric divides all correct predictions (TP + TN) by the total number of predictions.
What is accuracy?
In this popular clustering algorithm, you must tune the parameter "k," the number of clusters.
What is K-means?
This "flavor" of explainable AI uses a separate model to explain a black-box model's decision after it's been made.
What is a post-hoc explanation?
This is the approach a retailer would use to discover natural customer segments from purchase histories that have no prior grouping.
What is unsupervised learning (clustering)?
Unlike Ridge, this regularization method can shrink a coefficient all the way to zero, so it also performs feature selection.
What is Lasso regression?
Random forests are built using this ensemble strategy, whose name is short for "bootstrap aggregating."
What is bagging?
A model suffering from this has low error on the training set but high error on the test set, having learned the training data too well.
What is overfitting?
Under the EU AI Act, banned uses such as social scoring and biometric categorization fall into this risk tier.
What is unacceptable risk?
In the animal dataset, "number of legs," "weight," and "fin?" are all examples of these inputs to the model.
What are features?
This metric measures "the proportion of variance explained by the model," where roughly 0.9 is very good and below 0.0 means the model makes no sense.
What is R²?
Adding more of these layers, sitting between the input and output, lets a neural network learn more complex, non-linear patterns.
What are hidden layers?
This second category of unsupervised learning, exemplified by PCA, reduces the number of features in a dataset.
What is dimensionality reduction?
An reinforcement learning task is formally modelled using this framework, defined by states, actions, and rewards.
What is a Markov Decision Process (MDP)?
Rule-based systems, expert systems, and search algorithms belong to AI but fall outside this subset, where computers learn without being explicitly programmed.
What is machine learning?
This is the term for what has happened when a model scores R² = 0.95 on training data but only 0.61 on the test data.
What is overfitting?
This single metric balances precision and recall by combining them into one score.
What is the F1 score?
High bias causes underfitting, while a high amount of this causes overfitting.
What is variance?
On the performance-versus-interpretability chart, this model sits at the high-performance, low-interpretability end, the opposite corner from linear regression.
What is a neural network?