Remember Every Machine Learning Metric: 75 Flashcards on Algorithms, Loss Functions and Evaluation

Machine learning has a vocabulary problem: half of an interview or a midterm is checking that you can define terms precisely and pick the right metric. Deck KD8576 — Machine Learning: Algorithms & Metrics on Flip It is built for exactly that: 75 short, definition-first flashcards covering the whole pipeline, from labelled data to hyperparameter tuning. Here is what is on them, and where the deck stops short.

What the 75 cards cover

• Framing: Tom Mitchell's E/T/P definition, supervised vs unsupervised vs reinforcement vs semi-supervised learning, classification vs regression, clustering, dimensionality reduction, anomaly detection, transfer learning
• Data and fit: training / validation / test split, features and labels, feature engineering, feature scaling, normalization vs standardization, one-hot encoding, feature selection, curse of dimensionality, overfitting, underfitting, the bias-variance tradeoff
• Metrics: confusion matrix, accuracy, precision, recall, F1 score, Type I vs Type II errors, ROC curve, AUC, cross-validation, k-fold
• Algorithms: linear regression, logistic regression, decision trees, random forests, ensembles, bagging, boosting, k-means, k-NN, SVM, neural networks, deep learning, PCA
• Training mechanics: gradient descent, SGD, mini-batch, loss functions, MSE, cross-entropy, learning rate (and what happens when it is too high or too low), batch, epoch, early stopping, regularization, L1/Lasso, L2/Ridge, dropout, batch normalization, grid search, random search

Nine cards worth memorizing verbatim

Where the deck stops, and what to add

The title promises metrics, but the metric set is the classification starter pack. There is no specificity (even though recall is labelled "sensitivity", which implies its partner), no precision-recall curve, and no regression metrics at all — MSE only appears as a loss function, and MAE, RMSE and R² never show up. Write those five cards yourself before an exam; they are guaranteed items.

The bigger hole is class imbalance. Accuracy is given as a clean formula with no warning, yet the classic trap question is a 99%-accurate model on a dataset that is 99% negative. Add one card: "Why can accuracy be misleading?" → "On imbalanced data a trivial majority-class predictor scores high; use precision, recall, F1 or AUC instead."

Two backs also deserve a footnote. "Learning rate too low → may get stuck in local minima" is the folk version; the real cost is very slow training and stalling on plateaus and saddle points. And bagging specifically resamples with replacement (bootstrap) — the card says "random subsets", which is the weaker version.

How to drill it

Study the metrics block as one object, not nine cards. Draw a 2×2 confusion matrix with four invented numbers, then compute accuracy, precision, recall and F1 from it in one sitting. Repeat with different numbers until the formulas come out without the cards — that single exercise covers nine of them.

Deck KD8576 is in the Flip It library at flip-it.app. For the tooling side of the same workflow, see the NumPy flashcards and the pandas essentials study guide.