Answer: AI is the field of building systems that perform tasks normally needing human intelligence – recognising images, understanding language, making decisions, or predicting outcomes. Modern AI is mostly data-driven: systems learn patterns from examples rather than following hand-written rules.
Answer: They are nested. AI is the broad goal. Machine Learning is a subset where systems learn from data. Deep Learning is a subset of ML using multi-layer neural networks. Generative AI is a class of (mostly deep learning) models that create new content such as text, images, code or audio.
Answer: Narrow AI is built for a specific task—spam filtering, face recognition, or a chatbot. Artificial General Intelligence (AGI) would match humans across any intellectual task and transfer learning between domains. Every system deployed today, including large language models, is still considered narrow or at best broad-but-not-general.
Answer: Supervised learning (labelled data, e.g., predicting house prices), unsupervised learning (no labels, e.g., customer clustering), reinforcement learning (an agent learns by reward and penalty, e.g., game playing, robotics), and self-supervised learning (labels created from the data itself, e.g., predicting the next word—the basis of LLM pre-training).
Answer: A model made of layers of connected nodes (neurons). Each neuron takes weighted inputs, adds a bias and applies an activation function. Layers are stacked – input, hidden, output – and weights are adjusted during training via backpropagation so the network’s predictions get closer to the correct answers.
Answer: It adds non-linearity to a neuron’s output. Without it, any number of stacked layers collapses into a single linear transformation, and the network cannot learn complex patterns. Common choices: ReLU (hidden layers), sigmoid (binary probability), softmax (multi-class probability), and GELU (used in transformers).
Answer: After a forward pass produces a prediction, the loss (error) is calculated. Backpropagation uses the chain rule to compute how much each weight contributed to that error, moving backwards layer by layer. An optimizer such as gradient descent or Adam then nudges each weight in the direction that reduces the loss.
Answer: An optimization method that minimizes the loss by repeatedly moving weights a small step in the opposite direction of the gradient. The step size is the learning rate. Variants include batch, stochastic (one sample), and mini-batch gradient descent; Adam and AdamW are adaptive versions widely used today.
Answer: Overfitting is when a model memorises training data, including noise, and performs poorly on new data. Prevention: more data, data augmentation, regularisation (L1/L2, dropout), early stopping, simpler models, cross-validation and removing leaky or irrelevant features.
Answer: Computer Vision lets machines interpret images and video. Use cases: defect detection on manufacturing lines, medical image analysis, number-plate recognition, retail shelf monitoring, document OCR and face verification in banking KYC. Core tasks are classification, object detection, segmentation and tracking.
Answer: NLP enables machines to understand and generate human language. Tasks include sentiment analysis, named entity recognition, translation, summarisation, question answering and chatbots. Modern NLP is dominated by transformer-based models such as BERT (understanding) and GPT-style models (generation).
Answer: A Convolutional Neural Network uses filters that slide over an image to detect local patterns – edges, textures, shapes – building up to complex features in deeper layers. Pooling reduces size. CNNs are used in image classification, object detection (YOLO family) and medical imaging, though vision transformers now compete strongly.
Answer: RNNs process sequences one step at a time and carry a hidden state; LSTMs add gates to remember longer context and reduce vanishing gradients. Transformers replaced them because self-attention looks at all tokens at once, trains in parallel on GPUs and handles long-range dependencies far better.
Answer: A neural network architecture (introduced in “Attention Is All You Need”, 2017) built on self-attention. Each token weighs its relevance to every other token, allowing context-aware representations. It has encoder and decoder blocks; BERT uses the encoder, GPT-style LLMs use the decoder. It is the foundation of modern GenAI.
Answer: Training is when the model learns – weights are updated using large data and heavy compute. Inference is using the trained model to make predictions on new inputs. Training is done occasionally and is expensive; inference happens constantly in production, so latency and cost per request matter most there.
Answer: Reusing a model trained on a large general dataset as the starting point for a related task. For example, fine-tuning a pre-trained ResNet on a few thousand defect images, or adapting a pre-trained language model to legal text. It saves data, time and compute.
Answer: An agent takes actions in an environment and learns a policy that maximizes cumulative reward through trial and error. Examples: game-playing agents (AlphaGo), robot control, dynamic pricing, and RLHF – where human preference feedback is used as a reward to align LLMs.
Answer: Data quality usually matters more than model choice. Poor, biased, incomplete or mislabelled data produces poor models regardless of algorithm. Most project effort goes into collecting, cleaning, labelling, validating and governing data, and into making sure training data reflects real production conditions.
Answer: Accuracy, precision (of predicted positives, how many are right), recall (of actual positives, how many were found), F1-score (balance of precision and recall), ROC-AUC and the confusion matrix. For imbalanced problems like fraud, precision, recall and PR-AUC are more meaningful than accuracy.
Answer: Bias is systematic unfairness in outputs towards certain groups. It enters through unrepresentative training data, historical discrimination in labels, proxy features (e.g. postcode standing in for ethnicity) and design choices. It is mitigated through diverse data, fairness metrics, audits and human review.
Answer: Responsible AI is the practice of designing and deploying AI that is fair, transparent, explainable, private, secure, safe, reliable and accountable. In practice it means documented model cards, bias testing, human oversight, data consent, audit trails and clear ownership of AI decisions.
Answer: XAI makes model decisions understandable to humans – essential in banking, insurance and healthcare. Techniques: SHAP (feature contribution per prediction), LIME (local surrogate models), feature importance, partial dependence plots, counterfactual explanations and attention/saliency maps for deep models.
Answer: The EU AI Act classifies AI systems by risk (unacceptable, high, limited, minimal) with obligations phasing in from 2025 onwards. Others include GDPR for personal data, India’s DPDP Act 2023, sector rules from RBI and SEBI, and frameworks such as NIST AI RMF and ISO/IEC 42001 for AI management systems.
Answer: In deep networks, gradients shrink as they are propagated backwards, so early layers barely learn. Solutions: ReLU-family activations, residual (skip) connections, batch or layer normalisation, careful weight initialisation (He/Xavier) and gated architectures like LSTM.
Answer: For each token, attention creates Query, Key and Value vectors. The query is compared with every key (dot product, scaled and softmaxed) to get weights, which are used to combine the values. The result is a context-aware representation. Multi-head attention runs several of these in parallel to capture different relationships.
Answer: Embeddings are dense numeric vectors that represent words, sentences, images or users so that similar items sit close together in vector space. They power semantic search, recommendations, clustering and RAG retrieval. Similarity is usually measured with cosine similarity.
Answer: Bias is error from over-simple assumptions (underfitting); variance is error from sensitivity to training data (overfitting). Increasing model complexity lowers bias but raises variance. The goal is the sweet spot that minimises total error on unseen data, found through validation and regularisation.
Answer: Check if the problem is repetitive, pattern-based and has enough quality data; whether rules would do the job more cheaply; whether some error is acceptable; and whether the value outweighs build and run cost. Many problems are solved better with SQL, rules or process change than with a model.
Answer: MLOps applies DevOps practices to machine learning – versioning of data and models, automated training pipelines, CI/CD, model registry, deployment, monitoring for drift and retraining. Without it, models degrade silently in production and cannot be reproduced or audited.
Answer: Data drift is when input distributions change (e.g. customer age profile shifts). Concept drift is when the relationship between inputs and target changes (e.g. what counts as fraud changes). Both are detected with monitoring (PSI, KS tests, performance tracking) and handled with retraining.
Answer: Discriminative models learn the boundary between classes – P(y|x) – e.g. logistic regression, most classifiers. Generative models learn how data is produced – P(x) or P(x,y) – and can create new samples, e.g. GANs, VAEs, diffusion models and LLMs.
Answer: A GAN has a generator that creates fakes and a discriminator that tries to spot them; they improve by competing. Diffusion models learn to reverse a gradual noising process, generating images from noise step by step. Diffusion now dominates image and video generation because training is more stable.
Answer: Resampling (SMOTE, undersampling), class weights in the loss, threshold tuning, using the right metrics (recall, precision, PR-AUC), anomaly-detection approaches and collecting more minority-class data. Always split before resampling to avoid leakage.
Answer: Running AI models on devices close to the data – phones, cameras, factory sensors, cars – instead of the cloud. Benefits: low latency, privacy, offline operation and lower bandwidth cost. It relies on model compression: quantisation, pruning, distillation and runtimes like ONNX Runtime or TensorFlow Lite.
Answer: Quantisation stores weights in lower precision (e.g. INT8 or 4-bit instead of FP32) to cut memory and speed up inference with small accuracy loss. Distillation trains a small “student” model to mimic a large “teacher”, giving a compact model that retains much of the performance.
Answer: Models that process and relate more than one data type – text, image, audio, video – in a single system. Examples: describing an image, answering questions over a chart or scanned invoice, voice assistants that see the screen. Most frontier models (GPT, Claude, Gemini families) are now multimodal.
Answer: Models that process and relate more than one data type – text, image, audio, video – in a single system. Examples: describing an image, answering questions over a chart or scanned invoice, voice assistants that see the screen. Most frontier models (GPT, Claude, Gemini families) are now multimodal.
Answer: Define a baseline before launch, then track hard metrics (time saved, cost reduced, revenue uplift, error rate) and adoption metrics (active users, task completion). Subtract total cost – build, licences, inference, maintenance, change management. Pilot with a control group where possible.
Answer: Inaccurate or hallucinated outputs, bias, data leakage and privacy breaches, security attacks (prompt injection, data poisoning), IP and copyright exposure, regulatory non-compliance, over-reliance by users and vendor lock-in. Each needs controls, owners and monitoring.
Answer: Training a shared model across many devices or organisations without moving raw data. Each node trains locally and sends only weight updates to a central server for aggregation. Used in mobile keyboards and healthcare where data cannot leave the source for privacy reasons.
Answer: Placing humans at key points – labelling, reviewing low-confidence predictions, approving high-risk actions or correcting outputs that feed back into training. It improves accuracy and accountability, and is often a regulatory expectation for high-risk decisions such as credit or hiring.
Scenario-Based Questions
Answer: Fraud is likely under 1% of transactions, so predicting “not fraud” for everything gives 99% accuracy. Switch to recall, precision, and PR-AUC; apply class weights or SMOTE on the training split only; tune the decision threshold against the cost of a missed fraud versus a false alarm; and review top features with the fraud team.
Answer: Compare recent input distributions with training data (PSI, KS test) to check data drift; check whether labels or business rules changed (concept drift); verify upstream pipelines for schema or unit changes; review performance by segment. Then retrain on recent data and set up automated drift alerts and a retraining schedule.
Answer: Start with the problem, not the technology: analyse ticket data to find high-volume, repetitive categories. Options could be ticket classification and routing, suggested replies for agents, or a self-service assistant grounded in the knowledge base. Pilot one use case, define metrics (handle time, CSAT, deflection rate), keep humans in the loop and scale after proving value.
Answer: Pause automated decisions. Audit the training data for historical bias, check for proxy features (career gaps, certain colleges or hobbies), measure fairness metrics such as demographic parity and equal opportunity, retrain with debiasing, and add human review of rejections. Document everything, as hiring is high-risk under most AI regulation.
Answer: Use transfer learning from a pre-trained vision model, apply heavy data augmentation (rotation, flips, lighting changes), consider synthetic data, use active learning so experts label the most informative new images, and evaluate with stratified cross-validation. A vision-language model with few-shot prompting is also worth testing as a baseline.
Answer: Use SHAP or similar to show the top factors behind that specific decision (e.g. high debt-to-income, short credit history) and provide counterfactuals—what would have changed the outcome. Explanations must be in plain language, logged for audit, and consistent with regulator expectations on adverse action notices.
Answer: Quantise or distil the model, batch requests, cache frequent results, use a smaller model for easy cases and route only hard ones to the large model, right-size hardware (GPU vs CPU), use autoscaling and serverless for spiky load, and profile to find bottlenecks outside the model.
Answer: Common causes: data leakage in the test set, training-serving skew (features computed differently in production), different data distribution in production, preprocessing mismatch, or stale features. Fix by reproducing the production pipeline in validation, using time-based splits, and adding feature parity checks.
Answer: Build a global model (e.g. LightGBM with lag, rolling, calendar, price and promotion features) rather than 10,000 separate models; handle new products with similarity-based cold-start; evaluate with WAPE/MAPE by category; produce forecasts with uncertainty intervals; and automate weekly retraining with monitoring.
Answer: Be honest: AI will automate parts of the work – data preparation, first-draft reports, routine queries – but judgement, business context, stakeholder handling, and accountability stay with people. The realistic outcome is analysts doing more, higher-value work, provided they are upskilled to use AI tools well.