This section covers advanced ML concepts including deep learning, transfer learning, model deployment, and more. These topics are increasingly relevant in real-world AI applications and technical interviews.
- Subset of ML based on artificial neural networks with multiple layers.
- Learns hierarchical features directly from data.
- Requires large datasets and GPU/TPU support.
- What is the architecture of a neural network?
- What are activation functions? Name a few.
- What is the vanishing gradient problem?
- How is deep learning different from traditional machine learning?
Problem: Classify images of handwritten digits (e.g. MNIST).
Solution: Use a deep neural network (DNN) or convolutional neural network (CNN) to achieve high accuracy on image classification.
- Reusing a pre-trained model on a new, related task.
- Common in image and NLP tasks.
- Saves training time and improves performance with less data.
- What is transfer learning and when is it used?
- How do you fine-tune a pre-trained model?
- What is the difference between feature extraction and fine-tuning?
Problem: Classify plant diseases using a small labeled image dataset.
Solution: Use a pre-trained CNN (like ResNet or EfficientNet) and fine-tune the last few layers on the plant images.
- Making a trained model available for real-time predictions or batch inference.
- Common tools: Flask, FastAPI, Docker, AWS SageMaker, TensorFlow Serving.
- What are the steps in deploying a model?
- What’s the difference between batch and real-time inference?
- How do you monitor model performance after deployment?
Problem: Deploy a sentiment analysis model for a live chat support tool.
Solution: Build a REST API using FastAPI, containerize with Docker, and host on a cloud platform.
- Understanding how a model makes decisions.
- Tools: SHAP, LIME, feature importance plots.
- Why is interpretability important in ML?
- How do SHAP values work?
- What is the difference between global and local interpretability?
- The process of optimizing model parameters that are not learned from data (e.g., learning rate, number of trees).
- Techniques: Grid Search, Random Search, Bayesian Optimization
- What is the difference between parameters and hyperparameters?
- How does grid search compare to random search?
- What tools can automate hyperparameter tuning?
- Monitoring deployed models for performance decay.
- Detecting concept drift (change in data distribution) and data drift (change in input features).
- What is concept drift and how do you handle it?
- How can you set up monitoring for a deployed model?
| Topic | Tools & Techniques |
|---|---|
| Deep Learning | TensorFlow, PyTorch, Keras, CNN, RNN |
| Transfer Learning | ResNet, BERT, Hugging Face Transformers |
| Deployment | Flask, FastAPI, Docker, AWS SageMaker |
| Interpretability | SHAP, LIME, Feature Importance |
| Hyperparameter Tuning | Grid Search, Random Search, Optuna |
| Monitoring & Drift | MLflow, Evidently AI, Prometheus |
- How does a CNN differ from a fully connected network?
- When would you use transfer learning vs training from scratch?
- What are the challenges in deploying ML models in production?
- How do you handle models that degrade over time?
- What’s the difference between model explainability and interpretability?
Next Steps:
Check outbehavioral_questions.mdto prepare for common project-based and HR-style interview questions.