[NEW] Professional Machine Learning Engineer

Model Development & Training (30%): Algorithm selection and hyper‑parameter optimization; Model evaluation metrics and cross‑validation strategies; Deep learning architectures and transfer learning; Regularization, ensembling, and model interpretability.

Deployment & Scaling (25%): Containerization with Docker and orchestration with Kubernetes; Model serving patterns (batch, online, streaming); Scalable inference pipelines and latency optimization; CI/CD for ML models and infrastructure as code.

Monitoring, Ethics & Maintenance (20%): Model drift detection and automated retraining; Performance monitoring, logging, and alerting; Bias detection, fairness, and responsible AI practices; Security, privacy, and compliance considerations.

B) Use downsampling on the majority class until the dataset is perfectly balanced at 50/50, then train a standard logistic regression model.

C) Implement a cost-sensitive learning approach by assigning a higher class weight to the minority class during model training.

D) Duplicate the positive class records iteratively until the dataset ratio reaches 20/80.

E) Drop the fraud detection labels and use an unsupervised K-Means clustering approach to find outliers.

F) Rely solely on accuracy as your primary evaluation metric to ensure the model performs well globally.

Option B is incorrect because extreme downsampling of a 99.8% majority class throws away massive amounts of useful data, leading to a poorly generalized model.

Option C is correct because adjusting class weights allows the model to learn the importance of the minority class directly during the training phase without creating synthetic data or dropping real data, making it highly scalable and effective.

Option D is incorrect because simple oversampling by exact duplication often leads to severe overfitting on those specific duplicated examples.

Option E is incorrect because ignoring labeled data to use unsupervised clustering wastes the explicit ground-truth labels you already have.

Option F is incorrect because accuracy is a terrible metric for highly imbalanced data; a model predicting "no fraud" 100% of the time would achieve 99.8% accuracy while failing completely at its actual task.

B) Manual tuning based on trial and error

C) Random Search

D) Bayesian Optimization

E) Gradient Descent Optimization on the hyperparameters

F) Early Stopping with default hyperparameters

Option B is incorrect because manual tuning is inefficient, subjective, and highly unlikely to find the optimal configuration in a 12-dimensional space.

Option C is incorrect because while Random Search is better than Grid Search for high dimensions, it does not use information from previous runs to inform future parameter selections.

Option D is correct because Bayesian Optimization builds a probabilistic model of the objective function and uses it to select the most promising hyperparameters to evaluate next, making it highly sample-efficient for expensive model training.

Option E is incorrect because hyperparameters (like learning rate or layer count) are often non-differentiable, meaning standard gradient descent cannot be applied to optimize them.

Option F is incorrect because while early stopping prevents overfitting, relying purely on default hyperparameters will almost certainly result in a suboptimal model for a complex computer vision task.

B) Use a batch prediction pipeline running on a scheduled cron job every hour.

C) Deploy the model as a containerized microservice on Kubernetes with Horizontal Pod Autoscaling (HPA) and cluster autoscaling enabled.

D) Serve the model directly from a cloud storage bucket using a static website hosting configuration.

E) Write the model weights into a relational database and perform inference using SQL queries.

F) Set up an edge deployment where the full NLP model is downloaded directly to the user's mobile browser on every request.

Option B is incorrect because batch prediction cannot serve real-time consumer API requests with sub-200ms latency; it is designed for offline, scheduled processing.

Option C is correct because containerizing the model on Kubernetes allows it to scale horizontally. HPA handles sudden traffic spikes by adding more pods, and it can scale down during low-traffic periods to reduce infrastructure costs.

Option D is incorrect because static storage buckets can serve static assets (like HTML or images) but cannot compute ML model inferences.

Option E is incorrect because relational databases sql complete guide for developers are not designed to load ML model weights and perform complex NLP inferences via SQL.

Option F is incorrect because downloading a heavy NLP model to a mobile browser on every request would cause massive latency, burn user bandwidth, and compromise the proprietary model weights.

You can retake the exams as many times as you want

This is a huge original question bank

You get support from me if you have questions

Each question has a detailed explanation

Mobile-compatible with the Udemy app