Stanford Machine Learning Practice Test

The specialization is broader because it goes beyond basic supervised learning into unsupervised methods, recommender systems, tree-based models, and neural networks. The official program description presents it as a wider foundational path across several core machine learning areas.

The learning rate alpha is too large, causing the parameter updates to overshoot the minimum; decrease the learning rate. is correct because If the learning rate alpha is set too large, each step of gradient descent will overshoot the minimum and land on a point on the opposite side of the cost valley that is higher than the starting point. The cited source, Machine Learning Specialization - Gradient Descent in Practice, supports this answer for the Supervised Learning scenario rather than the adjacent distractors.

The test set has to stay untouched so it can remain the final unbiased estimate of generalization. Once a team starts tuning against it, the test set stops acting like an independent check on the chosen model.

Different random starts can lead k-means to different local solutions. That is why practitioners commonly rerun the algorithm and compare outcomes instead of trusting a single initialization.

This is the vanishing gradient problem. The derivative of the Sigmoid function g'(z) is g(z)*(1-g(z)), which has a maximum value of only 0.25 (at z=0) and approaches 0 as z becomes highly positive or negative. During backpropagation, the gradient for early layers is calculated by multiplying the derivatives of many sigmoid activations. Multiplying many values that are less than 0.25 causes the gradient to shrink exponentially as it travels backward. Consequently, the weights in the early layers receive microscopic updates, causing training to stall.

Random Forests are based on bagging: they train multiple deep trees independently in parallel, and average their predictions to reduce variance. Gradient boosting (e.g., XGBoost) trains trees sequentially. Each new tree is typically shallow (low capacity) and is trained to predict the remaining residual errors (gradients of the loss function) made by the ensemble of all previously trained trees. This sequentially reduces the bias of the ensemble, building a highly powerful predictor step-by-step.

What current error analysis says about the model's main failure mode is correct because For the certification exam, this objective expects you to connect Practical ML Workflow to the specific task in the prompt: "A product manager asks whether the team should collect more examples or just keep tuning hyperparameters. The cited source, CS229: Machine Learning syllabus, supports this answer for the Practical ML Workflow scenario rather than the adjacent distractors.

Recommender systems are designed for exactly this job: predicting or ranking items a user is likely to prefer based on interaction patterns. The specialization treats this as a major applied machine learning use case.

It also covers unsupervised learning, recommender systems, tree-based models, and neural networks is correct because The specialization is broader because it goes beyond basic supervised learning into unsupervised methods, recommender systems, tree-based models, and neural networks. The cited source, Machine Learning Specialization, supports this answer for the Specialization Scope scenario rather than the adjacent distractors.

Regression fits this case because the team is predicting a continuous numeric value rather than assigning a label. The CS229 supervised-learning notes distinguish regression from classification by the type of output the model is asked to produce.