Machine Learning Engineer Interview Questions India 2026

Tip: Parameters are learned from data during training (weights, biases). Hyperparameters are set before training and control the learning process (learning rate, number of layers, batch size). You tune hyperparameters with cross-validation; you do not tune parameters directly.

Tip: Transfer learning uses a model pre-trained on a large dataset as a starting point. Most beneficial when: labelled data is scarce, compute budget is limited, or domains are similar (ImageNet to medical imaging). Fine-tune the last layers; freeze early layers.

Tip: Backpropagation computes the gradient of the loss function with respect to each weight using the chain rule. It propagates error signal from output layer backwards. The gradient tells the optimiser (SGD/Adam) how to adjust each weight to reduce loss.

Tip: In deep networks with sigmoid/tanh activations, gradients shrink exponentially during backprop — early layers learn very slowly. Solutions: ReLU activations, residual connections (ResNet skip connections), batch normalisation, gradient clipping for RNNs.

Tip: Data drift: input distribution shifts over time. Concept drift: the relationship between inputs and output changes. Detect with: monitoring prediction score distributions, input feature statistics, and business KPIs. Handle with: scheduled retraining, online learning.

Tip: Serve via REST API (FastAPI + uvicorn), containerise with Docker, orchestrate with Kubernetes. For low-latency: ONNX runtime or TensorRT. Use a model registry (MLflow) for versioning. A/B test new models via traffic splitting. Monitor prediction latency and drift.

Tip: Batch: run inference on a large dataset offline, results stored (e.g. daily churn predictions). Real-time: single-sample inference on demand (e.g. fraud detection during a transaction). Batch: higher throughput, cheaper. Real-time: low-latency requirement, more infrastructure complexity.

Tip: Cover all stages: problem framing, data collection/cleaning, feature engineering, model selection, training, evaluation, deployment, monitoring. Highlight one non-trivial decision at each stage. Show you understand the full MLOps lifecycle, not just modelling.

Tip: Training-serving skew: test data distribution differs from production. Target leakage: a feature in training was not available in production. Check feature distributions at serving time vs training time. Log and analyse production inputs.

Tip: Model versioning: MLflow, DVC, or Weights & Biases — track hyperparameters, metrics, artifacts per experiment. Dataset versioning: DVC or cloud storage with immutable versioned paths. Never overwrite a dataset used in training.

Tip: Validation set: used during development to compare models and tune hyperparameters. Test set: held out completely, used ONCE at the very end. Tuning on test data leaks information — the test accuracy then overestimates true generalisation.

Tip: PyTorch: Pythonic, easier debugging, dominant in research and increasingly in production. TensorFlow: mature serving stack (TF Serving, TFLite), better mobile/embedded support. Know one deeply rather than both superficially.