深層学習エンジニアリング

正解: To measure prediction error

A loss function quantifies the difference between predicted values and true values, guiding the optimization process. The stem asks: What is the main purpose of a loss function in machine learning?. The keyed answer is 1. Options: To measure prediction error; To update weights directly; To preprocess data; To reduce model size. Keep this in the MachineLearning bucket, and reject the neighbor term that solves a different job.

正解: Model fits training data too well and generalizes poorly

Overfitting occurs when a model captures noise instead of general patterns, reducing performance on unseen data. The stem asks: What does overfitting mean?. The keyed answer is 1. Options: Model fits training data too well and generalizes poorly; Model is too simple; Model cannot learn; Training data is insufficient. Keep this in the MachineLearning bucket, and reject the neighbor term that solves a different job.

正解: To prevent overfitting

Regularization adds constraints or penalties to reduce model complexity and improve generalization. The stem asks: What is the role of regularization?. The keyed answer is 1. Options: To prevent overfitting; To increase model complexity; To speed up inference; To normalize data. Keep this in the MachineLearning bucket, and reject the neighbor term that solves a different job.

正解: Logistic Regression

Logistic regression models the probability of class membership and is widely used for classification tasks. The stem asks: Which algorithm is mainly used for classification?. The keyed answer is 1. Options: Logistic Regression; Linear Regression; K-means; PCA. Keep this in the MachineLearning bucket, and reject the neighbor term that solves a different job.

正解: Number of neighbors

k determines how many nearby data points are considered when making a prediction. The stem asks: What does k in k-NN represent?. The keyed answer is 1. Options: Number of neighbors; Number of clusters; Learning rate; Feature count. Keep this in the MachineLearning bucket, and reject the neighbor term that solves a different job.

正解: Maximize margin between classes

SVM seeks a decision boundary that maximizes the margin between different classes. The stem asks: What is the main idea of Support Vector Machines?. The keyed answer is 1. Options: Maximize margin between classes; Minimize data size; Increase depth; Reduce features. Keep this in the MachineLearning bucket, and reject the neighbor term that solves a different job.

正解: Interpretability

Decision trees are easy to understand and visualize, making them highly interpretable. The stem asks: What is a key advantage of decision trees?. The keyed answer is 1. Options: Interpretability; High accuracy always; Low memory usage; Continuous optimization. Keep this in the MachineLearning bucket, and reject the neighbor term that solves a different job.

正解: Combine multiple models

Ensemble methods improve performance by aggregating predictions from multiple models. The stem asks: What does ensemble learning aim to do?. The keyed answer is 1. Options: Combine multiple models; Reduce dataset size; Eliminate features; Speed up labeling. Keep this in the MachineLearning bucket, and reject the neighbor term that solves a different job.

正解: Reducing variance

Bagging stabilizes predictions by training models on different bootstrap samples. The stem asks: What is bagging mainly used for?. The keyed answer is 1. Options: Reducing variance; Reducing bias; Feature selection; Dimensionality reduction. Keep this in the MachineLearning bucket, and reject the neighbor term that solves a different job.

正解: Bias and variance

Boosting focuses on hard examples to progressively improve model performance. The stem asks: What is boosting designed to improve?. The keyed answer is 1. Options: Bias and variance; Only variance; Only bias; Data quality. Keep this in the MachineLearning bucket, and reject the neighbor term that solves a different job.

正解: Loss function

Gradient descent iteratively updates parameters to minimize the loss function. The stem asks: What does gradient descent optimize?. The keyed answer is 1. Options: Loss function; Dataset size; Feature count; Model depth. Keep this in the MachineLearning bucket, and reject the neighbor term that solves a different job.

正解: Slow convergence

A very small learning rate makes training extremely slow. The stem asks: What happens if the learning rate is too small?. The keyed answer is 1. Options: Slow convergence; Divergence; Overfitting; No gradients. Keep this in the MachineLearning bucket, and reject the neighbor term that solves a different job.

正解: Stabilize optimization

Scaling ensures features contribute equally during optimization. The stem asks: What is the purpose of feature scaling?. The keyed answer is 1. Options: Stabilize optimization; Increase data size; Remove noise; Encode labels. Keep this in the Preprocessing bucket, and reject the neighbor term that solves a different job.

正解: Scaling values to a fixed range

Normalization typically rescales data into a range such as 0 to 1. The stem asks: What is normalization?. The keyed answer is 1. Options: Scaling values to a fixed range; Removing outliers; Encoding categories; Reducing dimensions. Keep this in the Preprocessing bucket, and reject the neighbor term that solves a different job.

正解: Mean 0 and variance 1

Standardization centers data and scales it by standard deviation. The stem asks: What is standardization?. The keyed answer is 1. Options: Mean 0 and variance 1; 0 to 1 scaling; Log transformation; Discretization. Keep this in the Preprocessing bucket, and reject the neighbor term that solves a different job.

正解: Dimensionality reduction

PCA reduces dimensions while preserving maximum variance. The stem asks: What is PCA mainly used for?. The keyed answer is 1. Options: Dimensionality reduction; Classification; Regression; Clustering. Keep this in the MachineLearning bucket, and reject the neighbor term that solves a different job.

正解: Curse of dimensionality

High dimensions increase sparsity and degrade model performance. The stem asks: What is the main risk of high-dimensional data?. The keyed answer is 1. Options: Curse of dimensionality; Over-regularization; Underfitting; Label noise. Keep this in the MachineLearning bucket, and reject the neighbor term that solves a different job.

正解: Prediction results by class

It summarizes correct and incorrect predictions for each class. The stem asks: What does a confusion matrix show?. The keyed answer is 1. Options: Prediction results by class; Loss curve; Feature importance; Model structure. Keep this in the MachineLearning bucket, and reject the neighbor term that solves a different job.

正解: Correct positive predictions ratio

Precision evaluates how many predicted positives are actually correct. The stem asks: What does precision measure?. The keyed answer is 1. Options: Correct positive predictions ratio; All correct predictions; Error rate; Recall. Keep this in the MachineLearning bucket, and reject the neighbor term that solves a different job.

正解: Detected positives ratio

Recall measures how many actual positives were correctly identified. The stem asks: What does recall measure?. The keyed answer is 1. Options: Detected positives ratio; Prediction confidence; Accuracy; Loss. Keep this in the MachineLearning bucket, and reject the neighbor term that solves a different job.