Why is Python preferred for AI development? Python is best suited for AI development as it has simple syntax, large libraries like NumPy, Pandas, Scikit-learn, TensorFlow, PyTorch, and strong community support. 1.

What are Python data types commonly used in AI? Python Data types commonly used in AI are Lists, Tuples, Dictionaries, Sets, NumPy arrays, Pandas DataFrames. 1.

What is Python’s Global Interpreter Lock (GIL)? Python’s Global Interpreter Lock is a mutex that allows only one thread to execute Python bytecode at a time 1.

What are Python decorators, Python comprehensions and Python generators? Python decorators: Functions that modify other functions or methods without changing their code.

Python comprehensions: one-line constructs for creating lists, dicts, or sets efficiently.

Python generators: Functions that yield values one at a time using yield, useful for handling large datasets.

Difference between AI, ML, and Deep Learning? AI is the concept of machines acting intelligent, ML is learning from data and Deep Learning is ML using neural networks. 1.

What is overfitting in ML? How do you prevent overfitting? When a model performs well on training data but poorly on unseen data. To prevent overfitting, use Regularization, dropout, early stopping and cross-validation. 1.

What is the bias-variance tradeoff? Low bias = complex model, low variance = simple model. The tradeoff is finding balance for generalization. 1.

What is feature engineering? Difference between classification and regression? Transforming raw data into meaningful features for ML models. Classification predicts categories and regression predicts continuous values. 1.

What are confusion matrix metrics? Confusion matrix metrics: Accuracy, Precision, Recall, F1-score. 1.

What is dimensionality reduction? Dimensionality reduction: Reducing input features while retaining key information (e.g., PCA, t-SNE). 1.

What are NumPy and pandas used for in AI? NumPy: Fast array operations, linear algebra, and numerical computations.

Pandas: Data manipulation and analysis using DataFrames.

What is Scikit-learn used for? Scikit-learn is used for ML algorithms (classification, regression, clustering, preprocessing). 1.

What are TensorFlow and PyTorch? TensorFlow: An open-source ML/DL library for building and training neural networks.

PyTorch: A deep learning framework with dynamic computation graphs.

Difference between TensorFlow and PyTorch? TensorFlow uses static graphs where PyTorch uses dynamic (eager execution). 1.

What are Keras and Hugging Face Transformers? Keras: A high-level API for building deep learning models.

Hugging Face Transformers: A library for state-of-the-art NLP models.

1. What are neural networks and perceptrons? Neural network: A system of interconnected nodes inspired by the human brain.

Perceptron: Simplest neural network unit with inputs, weights, bias, and activation function.

1. What are backpropagation and activation functions? Backpropagation: An algorithm for updating weights using gradient descent.

Activation function: A function that introduces non-linearity (ReLU, Sigmoid, Tanh, Softmax).

1. What are CNN (Convolutional Neural Network) and RNN (Recurrent Neural Network)? CNN: A deep learning model specialized for images.

RNN: A neural network for sequential data like text/time-series.

Difference between RNN, LSTM, and GRU? LSTM and GRU solve vanishing gradient problems better than vanilla RNNs. 1.

How do you monitor AI models in production? → Track accuracy, latency, drift, bias, fairness using monitoring tools (EvidentlyAI, Prometheus, Grafana). 1.

What is transfer learning and batch normalization? Transfer learning: Using pre-trained models for new tasks.

Batch normalization: Normalizing inputs within a layer to speed up training.

1. What is NLP and tokenization? Natural Language Processing: teaching machines to understand human language.

Tokenization: Splitting text into words or subwords.

Difference between image classification and object detection? Classification = one label per image, detection = identify and locate multiple objects. 1.

What are word embeddings and sentiment analysis Word embeddings: Vector representations of words (Word2Vec, GloVe, BERT embeddings).

Sentiment analysis: Classifying text into positive, negative, or neutral.

How do you handle imbalanced datasets? To handle imbalanced datasets, use oversampling, undersampling, SMOTE, weighted loss functions. 1.

What is named entity recognition (NER)? Named Entity Recognition→ Identifying entities like names, places, dates in text. 1.

What is computer vision and YOLO? Computer vision: AI field for processing and analyzing images/videos.

YOLO: “You Only Look Once” real-time object detection model.

How do you tune hyperparameters in ML models? Tune hyperparameters in ML models by grid search, Random search, Bayesian optimization. 1.

What is an adversarial attack in AI? How do you defend against adversarial attacks? Adversarial attack: Small perturbations to input data that fool ML models (common in image recognition).

To defend against adversarial attacks, use Adversarial training, defensive distillation, input preprocessing.

Difference between online learning and batch learning? Online = incremental learning, Batch = train on full dataset. 1.

What are embeddings in LLMs and explainable AI (XAI)? Embeddings: Dense vector representations of text tokens for semantic understanding.

Explainable AI (XAI): Making AI decisions interpretable (e.g., SHAP, LIME).

Difference between MLOps and DevOps? MLOps focuses on deploying, monitoring, and maintaining ML models in production. 1.

How do you deploy AI models? Deploy AI models by using Flask/FastAPI, Docker, Kubernetes, or cloud services (AWS SageMaker, Azure ML, GCP AI). 1.

How do you ensure reproducibility in ML experiments? Fix random seeds, use Docker environments, log experiments (MLflow, Weights & Biases). 1.

Why are Python metaclasses used in AI frameworks? Metaclasses define the behavior of classes themselves. AI libraries like PyTorch use metaclasses to implement dynamic graph creation and custom model classes. 1.

How do you optimize Python code for large-scale AI projects? To optimize python code for large-scale AI projects, use NumPy vectorization, multiprocessing, Cython, Numba JIT, or GPU acceleration with CuPy/PyTorch. 1.

How does Python handle concurrency for AI workloads? Python handles concurrency for AI workloads via:

Multiprocessing (bypasses GIL) Asyncio (for async I/O) GPU-based parallelism (CUDA, Tensor Cores)

What is memory profiling in Python, and how do you apply it in ML projects? Memory profiling analyzes RAM usage and tools like memory_profiler, tracemalloc, and PyTorch’s torch.cuda.memory_summary() help debug ML training. 1.

Explain Python’s slots and why it matters in AI. slots restricts object attributes, reducing memory overhead in large-scale AI datasets. 1.

What is stochastic gradient descent (SGD), and why is it better than batch gradient descent? SGD updates weights per mini-batch, making it faster and less memory-intensive than full-batch descent. 1.

What are gradient clipping and exploding gradients? Exploding gradients occur in deep networks and clipping limits gradient values to stabilize training. 1.

What is the difference between Adam, RMSProp, and SGD optimizers? Adam combines momentum + adaptive learning rates, RMSProp adapts per parameter, SGD is simple but requires tuning. 1.

What is learning rate scheduling? Learning rate scheduling is dynamically adjusting learning rate (step decay, cosine annealing, warm restarts) to improve convergence. 1.

Difference between XGBoost, LightGBM, and CatBoost? All are gradient boosting frameworks. LightGBM is faster, CatBoost handles categorical features efficiently, XGBoost is widely used and robust. 1.

How do you perform feature selection in high-dimensional data? Using filter methods (correlation, chi-square), wrapper methods (RFE), or embedded methods (Lasso, Tree-based importance). 1.

What is a Transformer model, and how does it improve over RNNs? Transformers use self-attention, allowing parallel processing and better long-term dependency handling vs sequential RNNs. 1.

Explain fine-tuning vs prompt-tuning vs parameter-efficient tuning. Fine-tuning updates all model weights; prompt-tuning adds task-specific tokens; parameter-efficient tuning modifies small adapter layers only. 1.

What is knowledge distillation in AI models and is quantization in deep learning? Training a smaller student model to mimic a larger teacher model’s predictions.

Quantization: Reducing model precision (FP32 → INT8) for faster inference with minimal accuracy loss.

Explain contrastive learning. → A self-supervised learning technique that learns embeddings by maximizing similarity of related pairs and minimizing unrelated pairs. 1.

How do you deploy large AI models efficiently? → Use model compression, quantization, ONNX export, GPU inference servers (TensorRT, TorchServe). 1.

What is Python’s role in AI? Python plays a central role in AI because it: 1.

Which Python packages are used in AI? Python AI relies heavily on packages like NumPy, Pandas, Scikit-learn, TensorFlow, PyTorch, Keras, NLTK, SpaCy, OpenCV, Hugging Face Transformers, and MLflow.