Basic Artificial Intelligence interview questions

150 questions with answers (AIE-001 to AIE-150). Try each one aloud before reading the answer.

Math & Statistics for ML

AIE-001 · What are vectors and matrices used for in machine learning?

A vector holds one sample's features or one set of weights; a matrix holds a whole dataset (rows = samples, columns = features) or a layer's weights. Most ML computation is matrix multiplication, which GPUs accelerate.

AIE-002 · What is a dot product, and why is it important in ML?

a·b = Σ aᵢbᵢ = ‖a‖‖b‖cos θ. It measures alignment between vectors and is the core operation of linear models, neuron pre-activations (w·x + b) and attention scores.

AIE-003 · What is a gradient?

The vector of partial derivatives of a function with respect to each input, ∇f = (∂f/∂x₁, …, ∂f/∂xₙ). It points in the direction of steepest increase, so optimizers step in the negative gradient direction to reduce loss.

AIE-004 · What is the chain rule, and why does it matter for neural networks?

For y = f(g(x)), dy/dx = f′(g(x))·g′(x). Backpropagation applies the chain rule layer by layer to compute the gradient of the loss with respect to every weight efficiently.

AIE-005 · What is the difference between mean, median and mode?

Mean is the arithmetic average, median is the middle value when sorted, and mode is the most frequent value. The median is robust to outliers; the mean is not.

AIE-006 · What are variance and standard deviation?

Variance σ² = E[(X − μ)²] measures spread around the mean; standard deviation σ = √variance is in the same units as the data, which makes it easier to interpret.

AIE-007 · What is the normal (Gaussian) distribution?

A symmetric bell-shaped distribution defined by mean μ and standard deviation σ. About 68%, 95% and 99.7% of values lie within 1σ, 2σ and 3σ of the mean. Many ML methods assume Gaussian noise.

AIE-008 · State Bayes' theorem.

P(A|B) = P(B|A)·P(A) / P(B). It updates a prior belief P(A) into a posterior P(A|B) after observing evidence B, and underlies Naive Bayes and Bayesian inference.

AIE-009 · What is conditional probability, and when are two events independent?

P(A|B) = P(A ∩ B)/P(B) is the probability of A given B occurred. A and B are independent when P(A ∩ B) = P(A)·P(B), i.e., P(A|B) = P(A).

AIE-010 · What are eigenvalues and eigenvectors?

For a square matrix A, an eigenvector v satisfies Av = λv: A only scales v by the eigenvalue λ. In PCA, eigenvectors of the covariance matrix give principal directions and eigenvalues give variance along them.

AIE-011 · What is a convex function, and why is convexity useful in optimization?

A function where the line segment between any two points on its graph lies on or above the graph. Any local minimum is a global minimum, so linear and logistic regression losses can be optimized reliably.

AIE-012 · What is the difference between covariance and correlation?

Covariance measures how two variables vary together but depends on their units. Pearson correlation divides covariance by both standard deviations, giving a unitless value in [−1, 1] that measures linear association.

AIE-013 · What is a p-value?

The probability, assuming the null hypothesis is true, of observing data at least as extreme as what was observed. A small p-value (e.g., below a chosen α of 0.05) is evidence against the null hypothesis.

Classical Machine Learning

AIE-014 · What is the difference between AI, ML and deep learning?

AI is the broad goal of building machines that act intelligently. Machine learning is a subset of AI where systems learn patterns from data instead of explicit rules. Deep learning is a subset of ML that uses multi-layer neural networks.

AIE-015 · What is the difference between supervised, unsupervised and reinforcement learning?

Supervised learning trains on labeled input–output pairs. Unsupervised learning finds structure (clusters, low-dimensional representations) in unlabeled data. Reinforcement learning learns a policy by interacting with an environment and receiving rewards.

AIE-016 · What is the difference between classification and regression?

Classification predicts a discrete class label (spam or not spam). Regression predicts a continuous value (house price). Some models, such as decision trees, can do both.

AIE-017 · What is linear regression?

A model that predicts y = w·x + b, fitting weights by minimizing mean squared error between predictions and targets. It assumes a roughly linear relationship and is a standard baseline for regression.

AIE-018 · What is logistic regression?

A linear classifier that passes w·x + b through the sigmoid σ(z) = 1/(1 + e⁻ᶻ) to output a probability. It is trained by minimizing log loss (binary cross-entropy).

AIE-019 · What is a decision tree?

A model that recursively splits data on feature thresholds, forming a tree whose leaves give predictions. It is interpretable and handles non-linear relationships, but a single deep tree overfits easily.

AIE-020 · What is a random forest?

An ensemble of decision trees, each trained on a bootstrap sample with a random subset of features considered at each split. Predictions are averaged (regression) or voted (classification), which reduces variance.

AIE-021 · How does k-nearest neighbors (k-NN) work?

To predict a point, it finds the k closest training points under a distance metric and takes a majority vote or average. It has no training phase but is slow at prediction and sensitive to feature scaling.

AIE-022 · What is a support vector machine (SVM)?

A classifier that finds the hyperplane maximizing the margin between classes; the closest points are the support vectors. Kernels let it learn non-linear boundaries.

AIE-023 · What is k-means clustering?

An unsupervised algorithm that partitions data into k clusters by alternately assigning points to the nearest centroid and recomputing centroids as cluster means, minimizing within-cluster squared distance.

AIE-024 · What is Naive Bayes?

A probabilistic classifier that applies Bayes' theorem assuming features are conditionally independent given the class. It is fast, needs little data, and works well for text classification.

AIE-025 · What is the difference between bagging and boosting?

Bagging trains models independently on bootstrap samples and averages them, mainly reducing variance (random forest). Boosting trains models sequentially, each correcting the previous one's errors, mainly reducing bias (AdaBoost, gradient boosting).

AIE-026 · What is principal component analysis (PCA)?

An unsupervised linear technique that projects data onto orthogonal directions of maximum variance. Keeping the top components reduces dimensionality while preserving most of the variance.

Model Evaluation & Validation

AIE-027 · What is overfitting?

The model learns training data too closely, including noise, so it scores well on training data but poorly on unseen data. Signs are a large gap between training and validation performance.

AIE-028 · What is underfitting?

The model is too simple or under-trained to capture the underlying pattern, so it performs poorly on both training and validation data. Remedies include a more expressive model, better features or less regularization.

AIE-029 · What is a train/validation/test split?

The training set fits model parameters, the validation set guides hyperparameter tuning and model selection, and the held-out test set gives a final unbiased estimate of performance. The test set should be used only once.

AIE-030 · What is accuracy?

The fraction of predictions that are correct: (TP + TN) / total. It is misleading on imbalanced datasets, where always predicting the majority class can score highly.

AIE-031 · What is a confusion matrix?

A table of predicted versus actual classes. For binary problems it contains true positives, false positives, true negatives and false negatives, from which precision, recall and accuracy are computed.

AIE-032 · What is the difference between MSE and MAE?

MSE = mean of squared errors; it penalizes large errors heavily and is sensitive to outliers. MAE = mean of absolute errors; it is more robust to outliers and is in the target's units.

AIE-033 · What does R² measure?

The coefficient of determination: R² = 1 − SS_res/SS_tot, the fraction of target variance explained by the model. 1 is a perfect fit; 0 means no better than predicting the mean.

AIE-034 · What is an ROC curve?

A plot of true positive rate (recall) against false positive rate as the classification threshold varies. The area under it (ROC-AUC) is the probability that a random positive is scored above a random negative.

AIE-035 · Why should you build a baseline model first?

A simple baseline (majority class, mean prediction, linear model) sets a reference point. Any complex model must beat it to justify its cost, and it quickly exposes data or pipeline bugs.

AIE-036 · What is the difference between a parameter and a hyperparameter?

Parameters are learned from data during training (weights, biases). Hyperparameters are set before training and control the learning process (learning rate, tree depth, regularization strength) and are tuned on validation data.

AIE-037 · What is data leakage?

When information unavailable at prediction time, such as the target or test-set statistics, leaks into training. It inflates offline scores and causes the model to fail in production.

AIE-038 · What is log loss (cross-entropy loss)?

For binary classification, log loss = −[y·log p + (1 − y)·log(1 − p)], averaged over samples. It rewards well-calibrated probabilities and heavily penalizes confident wrong predictions.

AIE-039 · What is a learning curve?

A plot of training and validation performance versus training-set size (or epochs). A persistent gap indicates high variance; both curves plateauing at poor performance indicates high bias.

Feature Engineering & Data Preparation

AIE-040 · What is a feature?

An individual measurable input variable used by a model, such as age, pixel intensity or word count. Feature quality often matters more than model choice.

AIE-041 · What is feature engineering?

Creating, transforming or selecting input variables using domain knowledge to make patterns easier for a model to learn, for example ratios, aggregates, date parts or text counts.

AIE-042 · What is one-hot encoding?

Converting a categorical variable with k categories into k binary columns, exactly one of which is 1. It avoids imposing a false order on nominal categories.

AIE-043 · What is label (ordinal) encoding, and when is it appropriate?

Mapping each category to an integer. It suits ordinal categories with a natural order (low < medium < high) and tree models; for nominal categories with linear models it implies a false ordering.

AIE-044 · What is the difference between normalization and standardization?

Min-max normalization rescales to a fixed range: x′ = (x − min)/(max − min). Standardization (z-score) gives zero mean and unit variance: z = (x − μ)/σ. Both help distance-based and gradient-based models.

AIE-045 · How can you handle missing values?

Drop rows or columns with too many missing values, impute with mean, median or mode, use model-based imputation, or add a missing-indicator flag. Some models, such as gradient-boosted trees, handle missing values natively.

AIE-046 · What is an outlier, and how can you handle it?

A point far from the bulk of the data, detected with z-scores, IQR rules or isolation methods. Options are to verify and remove errors, cap (winsorize), transform (log) or use robust models and losses.

AIE-047 · What is feature selection?

Choosing a subset of relevant features to reduce overfitting, training cost and complexity. Methods are filter (correlation, mutual information), wrapper (recursive feature elimination) and embedded (L1 regularization, tree importance).

AIE-048 · What is class imbalance?

When one class greatly outnumbers others, as in fraud detection. Models may ignore the minority class, so use resampling, class weights and metrics such as precision, recall and PR-AUC.

AIE-049 · What is binning (discretization)?

Grouping a continuous variable into intervals, such as age bands. It can capture non-linear effects in linear models and reduce noise, but it discards within-bin information.

AIE-050 · What is data augmentation?

Creating modified copies of training samples, such as flipped or cropped images or paraphrased text, to enlarge the dataset and improve generalization without collecting new labels.

AIE-051 · What is data labeling, and why does label quality matter?

Assigning ground-truth targets to raw data, often by human annotators. Noisy or inconsistent labels cap achievable accuracy, so clear guidelines and inter-annotator agreement checks are essential.

AIE-052 · What is the curse of dimensionality?

As the number of features grows, data becomes sparse, distances become less meaningful and the data needed to generalize grows rapidly. It hurts k-NN and clustering most; dimensionality reduction helps.

Deep Learning Fundamentals

AIE-053 · What is an artificial neuron (perceptron)?

A unit that computes a weighted sum of inputs plus a bias, z = w·x + b, and applies an activation function, a = f(z). A single perceptron can only learn linearly separable functions.

AIE-054 · What is an activation function?

A non-linear function applied to a neuron's pre-activation, such as ReLU, sigmoid or tanh. Without it, stacked layers collapse into a single linear transformation.

AIE-055 · What is ReLU?

Rectified linear unit: f(x) = max(0, x). It is cheap, does not saturate for positive inputs and so eases vanishing gradients, but neurons can "die" if they always output zero.

AIE-056 · What is the difference between sigmoid and tanh?

Sigmoid maps to (0, 1) and is used for probabilities; tanh maps to (−1, 1) and is zero-centered. Both saturate for large |x|, causing vanishing gradients in deep networks.

AIE-057 · What is softmax?

A function that converts a vector of logits into a probability distribution: softmax(zᵢ) = e^zᵢ / Σⱼ e^zⱼ. It is used in the output layer for multi-class classification.

AIE-058 · What is backpropagation?

The algorithm that computes gradients of the loss with respect to all weights by applying the chain rule backward from the output layer, reusing intermediate results. An optimizer then updates the weights.

AIE-059 · What is the difference between an epoch, a batch and an iteration?

An epoch is one full pass over the training data. A batch is the subset of samples used for one gradient update. An iteration is one update, so iterations per epoch = dataset size / batch size.

AIE-060 · What is the learning rate?

The step size that scales each gradient update. Too high causes divergence or oscillation; too low makes training slow or stuck. It is usually the most important hyperparameter to tune.

AIE-061 · Which loss functions are commonly used for regression and classification?

Regression typically uses MSE, MAE or Huber loss. Binary classification uses binary cross-entropy; multi-class uses categorical cross-entropy with softmax.

AIE-062 · What is dropout?

A regularization technique that randomly zeroes a fraction p of activations during training, preventing co-adaptation of neurons. At inference all units are used, with activations scaled so expected values match.

AIE-063 · What is the Adam optimizer?

An adaptive optimizer that keeps running averages of gradients (first moment) and squared gradients (second moment), giving each parameter its own effective learning rate. It combines momentum with RMSProp-style scaling.

AIE-064 · What is a multilayer perceptron (MLP)?

A feed-forward network of fully connected layers: an input layer, one or more hidden layers with non-linear activations, and an output layer. It is the basic deep learning architecture for tabular data.

AIE-065 · What is batch normalization?

A layer that normalizes each feature over the mini-batch to zero mean and unit variance, then applies a learnable scale and shift. It stabilizes and speeds up training and allows higher learning rates.

CNNs & Computer Vision

AIE-066 · What is a convolutional neural network (CNN)?

A neural network that uses convolutional layers with shared, local filters to learn spatial features from grid data such as images. Stacked layers learn edges, then textures, then object parts.

AIE-067 · What is a convolution operation in a CNN?

Sliding a small filter over the input and computing the dot product at each position, producing a feature map. Deep learning libraries actually implement cross-correlation (no kernel flip), but call it convolution.

AIE-068 · What is a kernel (filter)?

A small learnable weight tensor, such as 3×3×C, that detects a specific local pattern. Each filter produces one output channel, and its weights are shared across all spatial positions.

AIE-069 · What are stride and padding?

Stride is how many pixels the filter moves per step; larger strides downsample. Padding adds border pixels (usually zeros) so the filter can cover edges and output size can be preserved ("same" padding).

AIE-070 · What is pooling?

A downsampling operation, such as max or average over a 2×2 window, that reduces spatial size and computation and adds some translation tolerance. It has no learnable parameters.

AIE-071 · What is a feature map?

The output of applying one filter across the input: a 2-D grid of activations showing where that filter's pattern appears. A conv layer outputs one feature map per filter.

AIE-072 · What is the difference between image classification, object detection and segmentation?

Classification assigns one label to the whole image. Detection locates objects with bounding boxes and labels. Segmentation labels every pixel, either by class (semantic) or by object instance (instance).

AIE-073 · What is transfer learning?

Reusing a model pretrained on a large dataset as the starting point for a new task, typically replacing the output head and fine-tuning. It needs far less labeled data and compute than training from scratch.

AIE-074 · What is a receptive field?

The region of the input image that influences a given activation. It grows with depth, kernel size, stride and pooling, letting deeper layers capture larger context.

AIE-075 · How is a color image represented as input to a CNN?

As a 3-D tensor of height × width × channels, typically 3 channels (RGB) with pixel values scaled to [0, 1] or standardized. A batch adds a fourth dimension.

AIE-076 · What is the role of fully connected layers in a CNN?

After convolutional feature extraction, fully connected (dense) layers, or a global pooling plus linear layer, combine features into final class scores. Global average pooling is often used to cut parameters.

AIE-077 · What is IoU (Intersection over Union)?

IoU = area of overlap / area of union between a predicted and a ground-truth box or mask. It ranges from 0 to 1 and decides whether a detection counts as correct at a given threshold, such as 0.5.

RNNs & Sequence Models

AIE-078 · What is a recurrent neural network (RNN)?

A neural network for sequences that processes one element at a time and keeps a hidden state that carries information from earlier steps. The same weights are reused at every time step, so it can handle inputs of varying length.

AIE-079 · What is the vanishing gradient problem in RNNs?

When gradients are backpropagated through many time steps, they are multiplied repeatedly by small values and shrink toward zero, so the network can't learn long-range dependencies. Exploding gradients are the opposite problem.

AIE-080 · What is an LSTM?

Long Short-Term Memory is an RNN variant with a cell state and input, forget and output gates that control what information is stored, forgotten and passed on. The gates let it learn dependencies over much longer sequences than a plain RNN.

AIE-081 · What is a GRU?

A Gated Recurrent Unit simplifies the LSTM with an update gate and a reset gate and no separate cell state. It has fewer parameters and often performs similarly to an LSTM.

AIE-082 · What is a sequence-to-sequence (seq2seq) model?

An encoder reads an input sequence into a representation, and a decoder generates an output sequence from it, possibly of different length. It is used for translation, summarisation and speech recognition.

AIE-083 · What is backpropagation through time (BPTT)?

Training an RNN by unrolling it across time steps into a deep feed-forward graph and applying backpropagation through all steps. Truncated BPTT limits how many steps are unrolled to save memory and time.

AIE-084 · What is gradient clipping?

Rescaling or capping gradients when their norm exceeds a threshold, so a single large update doesn't destabilise training. It is a standard fix for exploding gradients in RNNs.

AIE-085 · What is a bidirectional RNN?

It runs one RNN forward and another backward over the sequence and combines their states, so each position sees both past and future context. It is useful when the whole sequence is available, such as tagging words in a sentence.

AIE-086 · What is teacher forcing?

During training of a sequence generator, the true previous token is fed as the next input instead of the model's own prediction. It speeds training but can cause a mismatch at inference, called exposure bias.

AIE-087 · What is time-series forecasting?

Predicting future values of a sequence from its past values and related variables, such as sales, demand or sensor readings. Models range from ARIMA to gradient boosting, RNNs and transformers.

AIE-088 · What is a word embedding?

A dense vector representation of a word learned so that words used in similar contexts have similar vectors. Word2Vec and GloVe are classic methods; transformers learn contextual embeddings.

AIE-089 · What is tokenisation in NLP?

Splitting text into units (tokens) such as words, subwords or characters that a model processes. Subword methods like BPE or WordPiece balance vocabulary size with handling rare words.

AIE-090 · What is the difference between many-to-one and many-to-many sequence tasks?

Many-to-one maps a whole sequence to a single output, such as sentiment classification of a review. Many-to-many maps a sequence to a sequence, such as part-of-speech tagging or translation.

Transformers & NLP

AIE-091 · What is a transformer?

A neural network architecture based on self-attention instead of recurrence, which processes all tokens in parallel. It is the foundation of modern language models and is widely used in vision and speech.

AIE-092 · What is self-attention?

Each token computes how much to attend to every other token in the sequence using query, key and value vectors, then forms a weighted sum of values. It lets the model capture relationships regardless of distance.

AIE-093 · What are queries, keys and values in attention?

Each token is projected into a query (what it is looking for), a key (what it offers) and a value (the information it carries). Attention weights come from query–key similarity, and the output is the weighted sum of values.

AIE-094 · What is multi-head attention?

Several attention operations run in parallel with different learned projections, so each head can focus on different relationships (syntax, position, meaning). Their outputs are concatenated and projected.

AIE-095 · What is positional encoding?

Self-attention has no built-in notion of order, so position information is added to token embeddings, using fixed sinusoids, learned position embeddings or relative schemes such as RoPE.

AIE-096 · What is the difference between encoder-only, decoder-only and encoder-decoder transformers?

Encoder-only models (BERT) build bidirectional representations for understanding tasks; decoder-only models (GPT) generate text left to right; encoder-decoder models (T5) map an input sequence to an output sequence, as in translation.

AIE-097 · What is BERT?

Bidirectional Encoder Representations from Transformers, an encoder model pretrained with masked language modelling. It is fine-tuned for tasks like classification, question answering and named entity recognition.

AIE-098 · What is masked language modelling?

A pretraining task where some input tokens are hidden and the model learns to predict them from surrounding context. It teaches bidirectional understanding.

AIE-099 · What is causal (autoregressive) language modelling?

The model predicts the next token given only previous tokens, using a causal mask so it can't look ahead. It is how GPT-style models are trained and how they generate text.

AIE-100 · What is named entity recognition (NER)?

An NLP task that finds and labels entities in text, such as people, organisations, locations and dates. It is usually framed as token classification.

AIE-101 · What is sentiment analysis?

Classifying text by the opinion or emotion it expresses, such as positive, negative or neutral. It is used for reviews, social media and customer feedback.

AIE-102 · What is a context window?

The maximum number of tokens a transformer can attend to at once. Text longer than the window must be truncated, chunked or summarised.

LLMs, Prompting, RAG & Agents

AIE-103 · What is a large language model (LLM)?

A very large transformer, usually decoder-only, pretrained on massive text corpora to predict the next token. With instruction tuning it can follow instructions, answer questions, write and reason across many tasks.

AIE-104 · What is prompt engineering?

Designing the input instructions, context and examples given to an LLM to get accurate, well-formatted outputs. It includes clear instructions, role setting, examples and output format constraints.

AIE-105 · What is the difference between zero-shot and few-shot prompting?

Zero-shot gives only instructions; few-shot also includes a few worked examples in the prompt. Examples help the model infer format and task details without retraining.

AIE-106 · What is chain-of-thought prompting?

Asking the model to work through intermediate reasoning steps before the answer. It often improves accuracy on multi-step problems such as maths and logic.

AIE-107 · What is a hallucination in an LLM?

A fluent, confident output that is factually wrong or unsupported by the source material. It arises because models generate plausible text rather than looking up verified facts.

AIE-108 · What is retrieval-augmented generation (RAG)?

A pattern where relevant documents are retrieved from a knowledge source (often a vector database) and added to the prompt, so the LLM answers from that context. It improves factual accuracy and lets the model use private or up-to-date information without retraining.

AIE-109 · What is a vector database?

A store that indexes embedding vectors and supports fast similarity search (nearest neighbours), usually with approximate algorithms like HNSW. It is the retrieval engine behind semantic search and RAG.

AIE-110 · What is an embedding in the context of LLM applications?

A numeric vector representing the meaning of text (or images), produced by an embedding model, such that similar meanings have nearby vectors. Embeddings power semantic search, clustering and retrieval.

AIE-111 · What is fine-tuning an LLM?

Continuing training of a pretrained model on task- or domain-specific examples so it follows a format, style or specialised knowledge better. Parameter-efficient methods like LoRA update only small adapter weights.

AIE-112 · What is an AI agent?

An LLM-driven system that plans steps, calls tools (search, code, APIs), observes results and iterates toward a goal, rather than producing a single response.

AIE-113 · What is temperature in text generation?

A sampling parameter that sharpens or flattens the next-token probability distribution. Low temperature gives more deterministic, focused outputs; high temperature gives more varied, creative and error-prone outputs.

AIE-114 · What is a token in the context of LLMs?

The unit of text the model reads and writes, often a word piece of a few characters. Context limits, pricing and speed are all measured in tokens.

Generative Models (GANs, VAEs, Diffusion)

AIE-115 · What is a generative model?

A model that learns the probability distribution of data so it can create new samples resembling the training data, such as images, text or audio.

AIE-116 · What is a GAN?

A Generative Adversarial Network has a generator that creates fake samples and a discriminator that tries to tell real from fake. Trained against each other, the generator learns to produce realistic data.

AIE-117 · What is a variational autoencoder (VAE)?

An autoencoder that encodes inputs into a probability distribution in latent space and decodes samples from it. A KL-divergence term keeps the latent space smooth so new data can be generated by sampling.

AIE-118 · What is an autoencoder?

A network that compresses input into a smaller latent representation (encoder) and reconstructs it (decoder). It is used for dimensionality reduction, denoising and anomaly detection.

AIE-119 · What is a diffusion model?

A model trained to reverse a gradual noising process: it learns to remove noise step by step, so it can generate data by starting from pure noise. It powers modern image generators.

AIE-120 · What is latent space?

A compressed representation space learned by a model, where similar inputs lie close together. Moving through latent space can smoothly change generated outputs.

AIE-121 · What is mode collapse in GANs?

The generator produces only a few kinds of outputs instead of the full variety in the data, because it finds a narrow set that fools the discriminator.

AIE-122 · What is text-to-image generation?

Generating images from text prompts, typically with diffusion models conditioned on text embeddings from a language or vision-language encoder.

AIE-123 · What is the reparameterisation trick in a VAE?

Instead of sampling z directly from N(μ, σ²), the model samples ε from N(0, 1) and computes z = μ + σ·ε. This keeps the sampling step differentiable so gradients can flow to μ and σ during training.

AIE-124 · What is a deepfake?

Synthetic media in which a person's face, voice or actions are generated or altered by AI to look real. It raises concerns about misinformation, fraud and consent, driving work on detection and provenance.

AIE-125 · What is FID (Fréchet Inception Distance)?

A metric comparing the statistics of features from generated images and real images, extracted by an Inception network. Lower FID means generated images are closer to real ones in quality and diversity.

AIE-126 · What is conditional generation?

Generating samples that satisfy a given condition, such as a class label, text prompt or input image. Conditional GANs and text-guided diffusion models are examples.

Reinforcement Learning

AIE-127 · What is reinforcement learning?

Learning by trial and error: an agent takes actions in an environment, receives rewards, and learns a policy that maximises cumulative reward over time.

AIE-128 · What are the agent, environment, state, action and reward?

The agent makes decisions; the environment responds; the state describes the current situation; an action is the agent's choice; and the reward is the numeric feedback signal for that step.

AIE-129 · What is a policy?

A mapping from states to actions (or to probabilities of actions) that defines how the agent behaves. RL aims to find the policy that maximises expected return.

AIE-130 · What is a Markov decision process (MDP)?

A formal model of decision-making with states, actions, transition probabilities, rewards and a discount factor, where the next state depends only on the current state and action (Markov property).

AIE-131 · What is the exploration-exploitation trade-off?

The agent must balance trying new actions to discover better rewards (exploration) with choosing actions it already knows work well (exploitation). Epsilon-greedy and UCB are common strategies.

AIE-132 · What is a discount factor?

γ between 0 and 1 that reduces the value of future rewards; return = r₁ + γr₂ + γ²r₃ + …. Lower γ makes the agent short-sighted; higher γ values long-term reward.

AIE-133 · What is a value function?

V(s) estimates the expected return from a state when following a policy; Q(s, a) estimates the expected return of taking action a in state s and following the policy afterwards.

AIE-134 · What is Q-learning?

An off-policy algorithm that learns Q-values with the update Q(s,a) ← Q(s,a) + α[r + γ·max Q(s′,a′) − Q(s,a)], converging to optimal action values under suitable conditions.

AIE-135 · What is a deep Q-network (DQN)?

Q-learning with a neural network approximating Q(s, a), stabilised by experience replay and a separate target network. It learned to play Atari games from pixels.

AIE-136 · What is the difference between model-based and model-free RL?

Model-free methods learn policies or values directly from experience; model-based methods learn or use a model of the environment's dynamics to plan.

AIE-137 · What is reward shaping?

Adding extra rewards to guide learning toward desired behaviour faster. Poorly designed shaping can lead to unintended shortcuts.

AIE-138 · Where is reinforcement learning used in practice?

Game playing, robotics control, recommendation and ad bidding, resource scheduling, chip floorplanning, and aligning LLMs through RLHF.

MLOps, Deployment & Responsible AI

AIE-139 · What is MLOps?

Practices and tools for reliably building, deploying, monitoring and maintaining ML models in production, combining ML, DevOps and data engineering: versioning, automated pipelines, testing, deployment and monitoring.

AIE-140 · What is model deployment?

Making a trained model available to make predictions in a real application, for example behind a REST API, in a batch scoring job, on a mobile device or embedded in a stream-processing pipeline.

AIE-141 · What is the difference between batch and online (real-time) inference?

Batch inference scores large datasets on a schedule (e.g. nightly recommendations); online inference returns predictions per request with low latency (e.g. fraud checks at payment time).

AIE-142 · What is data drift?

A change in the distribution of input data in production compared with the training data, such as new customer behaviour or sensor changes. It can degrade model performance and should be monitored.

AIE-143 · What is concept drift?

A change in the relationship between inputs and the target, so the same inputs now lead to different outcomes (e.g. fraud patterns change). The model must be retrained or updated.

AIE-144 · What is a model registry?

A central store that tracks model versions, metadata, metrics, lineage and deployment stage (staging, production), supporting reproducibility and controlled releases.

AIE-145 · What is experiment tracking?

Recording parameters, code versions, data versions, metrics and artefacts of each training run (with tools such as MLflow or Weights & Biases) so results can be compared and reproduced.

AIE-146 · What is bias in machine learning models?

Systematic errors that produce unfair outcomes for certain groups, often caused by unrepresentative data, biased labels or proxy features. It is measured with fairness metrics across groups.

AIE-147 · What is explainability (XAI)?

Methods that help humans understand why a model made a prediction, such as feature importance, SHAP values, LIME and saliency maps. It supports trust, debugging and regulation.

AIE-148 · What is a model card?

A short document describing a model's intended use, training data, performance across conditions and groups, limitations and ethical considerations.

AIE-149 · What is model quantisation?

Reducing the numeric precision of model weights and activations (e.g. from 32-bit float to 8-bit integers) to cut memory, speed up inference and save power, usually with small accuracy loss.

AIE-150 · What is responsible AI?

Developing and using AI systems that are fair, transparent, accountable, privacy-preserving, safe and secure, with human oversight and compliance with laws and ethical guidelines.

Other levels: Intermediate · Advanced · Tricky · All Artificial Intelligence questions