Artificial Intelligence & Machine Learning MCQ

Explanation

ML systems learn patterns from training data to make predictions or decisions on new data. Instead of hand-coding rules, patterns are automatically discovered from examples.

Explanation

Supervised: spam detection (emails labeled spam/not). Unsupervised: clustering customers by purchasing behavior without predefined groups. Semi-supervised and reinforcement learning are other categories.

Explanation

Overfitting: high training accuracy, low test accuracy. The model memorizes training examples rather than learning generalizable patterns. Prevented by regularization, dropout, cross-validation, more data, or simpler models.

Explanation

Neural networks have input, hidden, and output layers. Each neuron applies a weighted sum + activation function. Deep neural networks (many hidden layers) learn hierarchical representations.

Explanation

Gradient descent computes the gradient of the loss with respect to parameters and updates them: θ = θ - α∇L. α is the learning rate. Variants: batch, stochastic (SGD), mini-batch.

Explanation

The test set must not be used during development to get an unbiased estimate. Using test data to make decisions contaminates it, creating optimistic biased estimates.

Explanation

Decision trees split at each node based on the feature best separating classes/reducing variance (Gini impurity, information gain, MSE). Easy to interpret but prone to overfitting without pruning.

Explanation

K-means: initialize k centroids, assign each point to nearest centroid, recompute centroids, repeat until convergence. Limitation: assumes spherical clusters, sensitive to initialization and outliers.

Explanation

Linear regression: y = w₀ + w₁x₁ + ... + wₙxₙ. Minimizes mean squared error. Assumes linear relationship, normal errors, and no multicollinearity. Ridge/Lasso add regularization.

Explanation

Logistic regression outputs P(y=1|x) = σ(wᵀx) where σ is sigmoid. Despite the name, it is a classification algorithm. Trained with cross-entropy loss. Interpretable coefficients as log-odds.

Explanation

Random Forests use bagging (bootstrap aggregation) and random feature subsets, reducing overfitting of individual trees. Aggregation: majority vote (classification) or average (regression). Feature importance is a useful output.

Explanation

High bias: model too simple, misses patterns (underfitting). High variance: model too complex, overfits. Total error = bias² + variance + irreducible noise. Regularization, ensemble methods, and more data help balance this.

Explanation

k-fold CV gives a more reliable estimate of generalization performance. Stratified k-fold preserves class distribution. Leave-one-out (LOO) uses each sample as the test set once.

Explanation

Precision: of all predicted positive, how many are actually positive (TP/(TP+FP)). Recall (sensitivity): of all actual positive, how many were found (TP/(TP+FN)). F1 = harmonic mean of both.

Explanation

F1 score is useful when classes are imbalanced and both false positives and false negatives matter. High F1 requires both high precision and high recall. F-β generalizes by weighting recall β times more than precision.

Explanation

Feature engineering includes normalization, encoding categoricals, creating interaction terms, log-transforming skewed features, and domain-specific transformations. Good features often matter more than model choice.

Explanation

SVM finds the maximum-margin hyperplane. Support vectors are the closest training points. Kernel trick (RBF, polynomial) maps data to higher dimensions for non-linear separation. Effective in high-dimensional spaces.

Explanation

Deep learning (CNNs, RNNs, Transformers) learns features automatically from raw data (pixels, text, audio). Enabled by: large datasets (ImageNet), GPUs, and algorithmic advances (ReLU, batch norm, attention).

Explanation

CNNs use convolutional layers (learn local feature detectors), pooling layers (spatial downsampling), and fully connected layers. Local connectivity and weight sharing make them efficient for images.

Explanation

RNNs process sequences by passing hidden state h_t = f(h_{t-1}, x_t) through time. Challenges: vanishing/exploding gradients. LSTMs and GRUs solve this with gating mechanisms.

Explanation

Pre-trained models (ImageNet-trained CNNs, BERT, GPT) capture general knowledge. Fine-tuning adapts them to specific tasks with small datasets. Reduces training time from days to hours.

Explanation

NLP tasks: sentiment analysis, named entity recognition, machine translation, question answering, text summarization. Transformers (BERT, GPT) have revolutionized NLP with self-attention mechanisms.

Explanation

RL components: agent, environment, state, action, reward. The agent learns a policy π(a|s) mapping states to actions. Applications: game playing (AlphaGo), robotics, recommendation systems, autonomous driving.

Explanation

L1 regularization (Lasso): adds λ|w| — promotes sparsity (many weights = 0). L2 regularization (Ridge): adds λw² — shrinks weights toward 0. Elastic Net combines both. Dropout is regularization for neural networks.

Explanation

Confusion matrix rows = actual class, columns = predicted class. TP: correct positives. TN: correct negatives. FP: false alarms. FN: missed detections. Derives precision, recall, F1, accuracy, and ROC curves.

Explanation

PCA finds lower-dimensional projections maximizing variance. t-SNE/UMAP are for visualization. Autoencoders learn compressed representations. Reduces overfitting, computational cost, and storage requirements.

Explanation

In high dimensions: distances between points converge, volume increases exponentially (more data needed), nearest neighbor becomes unreliable. Dimensionality reduction and feature selection help combat this.

Explanation

Discriminative (logistic regression, SVM, neural net classifiers): learn decision boundary P(Y|X). Generative (Naive Bayes, VAE, GAN, LLMs): model the data distribution, enabling generation of new samples.

Explanation

Data augmentation (for images: random crop, flip, color jitter; for text: back translation, synonym replacement) reduces overfitting and improves generalization when data is limited.

Explanation

ROC curve: TPR (sensitivity) vs FPR (1-specificity) at different thresholds. AUC=1: perfect. AUC=0.5: random. AUC measures ranking quality — probability that a random positive example is ranked higher than a random negative.

Explanation

Model parameters (weights) are learned from data. Hyperparameters (learning rate, batch size, number of trees, k in k-means) are set by the practitioner and tuned via grid search, random search, or Bayesian optimization.

Explanation

Batch normalization normalizes activations to have zero mean and unit variance within each mini-batch, then applies learned scale (γ) and shift (β). Stabilizes training, acts as regularizer, and allows faster convergence.

Explanation

Dropout: randomly set each neuron output to 0 with probability p during training. Forces the network to learn redundant representations. At inference, outputs are scaled by (1-p). Acts as ensemble of thinned networks.

Explanation

Attention (Bahdanau, 2014) allows models to focus on relevant input parts per output token. Self-attention (Transformers) computes relationships within the same sequence, enabling long-range dependency modeling.

Explanation

GPT models (OpenAI) use decoder-only Transformers trained on large text corpora via next-token prediction. Fine-tuning and RLHF align them for following instructions. GPT-4, Claude, Gemini are examples.

Explanation

Cross-entropy loss for classification: -Σy_i log(ŷ_i). MSE for regression: (y-ŷ)². Hinge loss for SVM. The gradient of the loss guides gradient descent updates to minimize prediction error.

Explanation

Without activation functions, neural networks are just linear transformations. ReLU (max(0,x)) is most common. Sigmoid (output 0-1) for binary output. Softmax for multi-class output. GELU, Swish used in Transformers.

Explanation

Classification: email spam/not, image = cat/dog/car. Regression: predict house price, temperature. Both are supervised. Logistic regression is classification despite the name. Some problems can be formulated as either.

Explanation

softmax(z_i) = exp(z_i) / Σ exp(z_j). Amplifies the largest logit and suppresses others. Used in the output layer for multi-class classification with cross-entropy loss.

Explanation

Knowledge graphs (Google Knowledge Graph, Wikidata, Freebase) represent facts as (subject, relation, object) triples. Used for question answering, recommendation, entity disambiguation, and grounding LLMs.

Explanation

Transformers (Vaswani et al., 2017) replaced RNNs for NLP. Self-attention O(n²) allows parallel processing (no sequential dependency). Scales well with data and compute. Foundation for BERT, GPT, T5, and most modern AI.

Explanation

BERT (encoder-only, masked LM + NSP) learns bidirectional context — useful for classification, NER, QA. GPT (decoder-only, autoregressive) predicts next token — useful for generation. T5 uses encoder-decoder.

Explanation

Fine-tuning: initialize with pre-trained weights, train on task data (often with a small learning rate). Works because pre-trained representations transfer well. PEFT methods (LoRA, Adapters) fine-tune efficiently with few parameters.

Explanation

GAN (Goodfellow 2014): generator G maximizes discriminator D's error; D minimizes it. At equilibrium, G generates realistic data. Used for image synthesis (StyleGAN), image-to-image translation (pix2pix), deepfakes.

Explanation

VAE encodes inputs to a distribution (μ, σ) rather than a point. Decoder samples from this distribution. KL divergence regularizes the latent space. Enables generation (sample from latent space) and smooth interpolation.

Explanation

Word2vec (Mikolov 2013) uses CBOW or skip-gram training to learn 100-300 dimensional word embeddings. Famous property: king - man + woman ≈ queen. Captured semantic relationships. Succeeded by contextual embeddings (BERT, GPT).

Explanation

Sigmoid/tanh saturate with small gradients (≤ 0.25). Multiplying many <1 values → 0. Solutions: ReLU activations, residual connections (ResNet), batch normalization, LSTM gates, gradient clipping, careful initialization.

Explanation

LSTM (Hochreiter 1997) cell state C_t has linear interaction — gradients flow back without vanishing. Gates control what to remember/forget/output. Handles long sequences (100-1000 steps), unlike vanilla RNNs.

Explanation

Object detection (YOLO, DETR, Faster R-CNN) outputs class + bounding box coordinates for each object. Image segmentation (semantic/instance) classifies every pixel. Classification: whole image → one label.

Explanation

ResNet (He et al., 2015): H(x) = F(x) + x. The skip connection allows gradients to bypass layers. This enabled training networks with 100-1000 layers, achieving breakthrough results on ImageNet.

Explanation

RLHF (InstructGPT, ChatGPT): collect human comparisons of outputs → train reward model → use PPO to maximize reward. Dramatically improves instruction following, helpfulness, and reducing harmful outputs.

Explanation

Prompt engineering techniques: zero-shot, few-shot (examples in prompt), chain-of-thought (reasoning steps), ReAct (reason+act). Crucial for getting desired behavior from LLMs without fine-tuning.

Explanation

RAG: retrieve relevant documents from a vector database using semantic search, prepend them to the prompt. The LLM conditions on retrieved context. Reduces hallucination, enables knowledge updates without retraining.

Explanation

With severe class imbalance, ROC AUC can be misleadingly optimistic (many TN inflate FPR denominator). PR AUC focuses on the minority positive class. For fraud detection (rare events), PR is more informative.

Explanation

Exploding gradients: gradient norm grows exponentially in deep/RNN networks. Gradient clipping: if ||g|| > threshold, g = g * threshold/||g||. Essential for training RNNs and very deep networks.

Explanation

Distillation (Hinton 2015): student learns from teacher's soft labels (temperature-scaled softmax) carrying inter-class similarity info. The student can approach teacher performance with far fewer parameters.

Explanation

NAS (AutoML): automates network design (layer types, connections, widths). EfficientNet, NASNet, MobileNetV3 were NAS-discovered. DARTS (differentiable NAS) makes search end-to-end differentiable and much faster.

Explanation

Collaborative filtering: find similar users (user-based) or items (item-based). Matrix factorization (SVD, ALS) decomposes user-item matrix. Deep learning (YouTube, Netflix): two-tower retrieval + ranking models.

Explanation

Concept drift: P(Y|X) changes over time (e.g., spam content evolves, economic conditions shift). Requires monitoring (data drift, prediction drift), retraining triggers, and online learning or periodic retraining.

Explanation

Federated learning (Google, Apple): models train locally on-device (privacy preserved), send gradients/weights to server for aggregation. Used in keyboard prediction, medical imaging (hospitals can't share patient data).

Explanation

Vector databases (Pinecone, Weaviate, Chroma, pgvector) store embeddings and enable semantic search. Used in RAG for document retrieval, recommendation systems, and image search with HNSW or IVF indexing.

Explanation

LoRA (Hu et al., 2021): instead of fine-tuning all weights W, add ΔW = BA where B (d×r), A (r×d), r<<d. Train only A,B (0.1% of parameters). Enables LLM fine-tuning on consumer GPUs.

Explanation

A model can be accurate overall but unfair: higher error rates for minority groups (demographic parity, equalized odds). Fairness metrics often conflict with each other and with accuracy. Crucial for hiring, lending, criminal justice AI.

Explanation

MSE = mean((y-ŷ)²): sensitive to outliers (squared amplification). MAE = mean(|y-ŷ|): more robust. Huber loss combines both: quadratic for small errors, linear for large. Choice depends on tolerance for outliers.

Explanation

LIME (local linear approximations), SHAP (Shapley values for feature attribution), attention visualization, saliency maps. Required by GDPR (right to explanation), healthcare, and financial regulations.

Explanation

Decision trees naturally capture non-linear interactions and handle categorical and numeric features without scaling, while logistic regression assumes a roughly linear decision boundary in feature space. The right choice depends on the data's structure and interpretability needs.

Explanation

The elbow method helps pick a reasonable k for k-means by plotting inertia (within-cluster variance) versus k. The point where adding more clusters yields diminishing returns ("the elbow") is a good candidate for k. The silhouette score is a complementary technique.

Explanation

Stratified sampling keeps the class ratio consistent across train, validation, and test sets, which is especially important for imbalanced datasets — otherwise a random split could leave a set with too few minority-class examples to evaluate fairly.

Explanation

Assigning integers like 1, 2, 3 to unordered categories implies an ordering or magnitude that does not exist. One-hot encoding creates a separate binary column per category, letting the model treat them as independent without artificial ordinal relationships.

Explanation

As depth increases, a tree can fit training data (and its noise) more closely, increasing variance and the risk of overfitting. Techniques like pruning, setting a maximum depth, or requiring a minimum samples per leaf help balance fit and generalization.

Explanation

A pipeline bundles steps like scaling, encoding, and the final estimator so that fit/transform are applied in the same order on both training and test data, which prevents accidentally leaking information from the test set into preprocessing statistics.

Explanation

Early stopping tracks a validation metric (such as validation loss) and stops training once it stagnates or worsens for a set number of epochs (patience). This acts as an implicit regularizer, halting training before the model starts memorizing the training set.

Explanation

Bagging (e.g., Random Forest) reduces variance by averaging independent models trained on bootstrapped subsets. Boosting (e.g., AdaBoost, Gradient Boosting, XGBoost) builds models sequentially, with each new model emphasizing the mistakes of the ensemble so far, typically reducing bias.

Explanation

Adam combines momentum (using a moving average of past gradients) with per-parameter adaptive learning rates (using a moving average of squared gradients), which often makes training more stable and faster to converge than vanilla SGD, especially on noisy or sparse gradients.

Explanation

Data leakage happens when features or preprocessing statistics are derived using information that would not be available at prediction time (such as fitting a scaler on the full dataset before splitting). The result is a model that looks great in evaluation but performs poorly in production.

Explanation

A learning rate schedule changes the step size used in gradient updates over time. A common pattern is to start with a larger learning rate to make rapid progress and gradually reduce it so the optimizer can settle into a minimum without overshooting.

Explanation

Tokenization converts text into discrete units — often subword pieces via algorithms like Byte-Pair Encoding (BPE) or WordPiece — which are then mapped to integer IDs and embedded. Subword tokenization balances vocabulary size with the ability to represent rare or unseen words.

Explanation

Learning curves reveal whether more data would help: a persistent gap between low training error and high validation error suggests overfitting (high variance), while both errors converging to a high value suggests the model is too simple (high bias) and needs more capacity or better features.

Explanation

Distance-based algorithms (k-NN, SVM, k-means) compute similarity using feature magnitudes. If one feature ranges from 0-1 and another from 0-100,000, the larger-scale feature will dominate the distance computation, so scaling features to comparable ranges (standardization or min-max scaling) is essential.

Explanation

Grid search becomes computationally expensive as the number of hyperparameters grows, since the number of combinations multiplies. Random search samples combinations randomly and often finds comparably good settings more efficiently, because typically only a few hyperparameters have a large effect on performance (Bergstra and Bengio, 2012).

Explanation

SSL (BERT: masked token prediction; GPT: next token prediction; CLIP: image-text contrastive; MAE: masked image patches) enables pre-training on internet-scale unlabeled data, learning powerful representations.

Explanation

Kaplan et al. (2020): loss ∝ N^-0.076 * D^-0.095 (N=params, D=tokens). Chinchilla (Hoffmann 2022): optimal: N and D should scale equally. These laws guide billion-dollar training run decisions.

Explanation

O(n²) attention quadruples cost when doubling sequence length. FlashAttention (IO-aware) keeps O(n²) compute but reduces memory bandwidth by computing in tiles. Sparse methods genuinely reduce compute for long sequences.

Explanation

MoE (Switch Transformer, GPT-4): each token routed to top-k experts (sparse activation). Total params >> active params per token. Enables scaling to hundreds of billions of parameters while keeping per-token compute manageable.

Explanation

Mechanistic interpretability (Anthropic, EleutherAI) identifies circuits: induction heads, copy suppression heads, indirect object identification circuits. Aims to fully understand transformer computations like reverse-engineering software.

Explanation

ICL (Brown et al., 2020): task examples in the prompt guide behavior without updating weights. Surprising emergent ability of large models. Fine-tuning updates weights: better at specific tasks but requires compute and can degrade other capabilities.

Explanation

CAI (Bai et al., 2022): SL-CAI critiques and revises its own outputs against a "constitution" of principles. RL-CAI uses AI feedback instead of human feedback. Scales alignment without proportional human labeling cost.

Explanation

Neural ODE (Chen 2018): instead of discrete layers, define dh/dt = f(h,t,θ). Solve with an ODE solver. Memory O(1) via adjoint method. Continuous normalizing flows, time-series modeling, and dynamical systems.

Explanation

Data parallelism (DDP, FSDP): scale to large datasets. Model parallelism (Tensor parallelism, pipeline parallelism): required when a model (e.g., 70B params) doesn't fit in a single GPU's memory. Megatron-LM uses both.

Explanation

Activation steering (Representation Engineering, Zou et al.): identify directions in activation space corresponding to concepts, add them during forward pass. Changes model behavior without fine-tuning. Related to mechanistic interpretability.

Explanation

Alignment: an AI system might optimize a proxy metric perfectly while causing unintended harm (Goodhart's law). RLHF, Constitutional AI, interpretability, and scalable oversight are approaches. Key concern for AGI safety research.

Explanation

Multi-modal models (GPT-4V, CLIP, Gemini, Flamingo) learn joint representations across modalities. CLIP learns image-text alignment via contrastive learning. Enables zero-shot image classification, image captioning, visual QA.

Explanation

Wei et al. (2022): abilities like chain-of-thought reasoning, arithmetic, and multi-step reasoning appear suddenly above certain model scales. Debate: are these genuinely emergent or artifacts of evaluation metrics? Has major implications for AI forecasting.

Explanation

Temperature T: logits/T before softmax. T=0 → greedy (always highest probability token). T=1 → sample from model distribution. T>1 → flatter distribution, more diverse but less coherent. Nucleus (top-p) and top-k sampling offer alternatives.

Explanation

Speculative decoding (Leviathan et al., 2022): draft model generates k tokens, target model verifies all in parallel (one forward pass). If draft tokens match target distribution, all are accepted. 2-3x speedup on typical text.

Explanation

RLHF: human raters compare outputs → reward model → PPO optimization. RLAIF: the AI itself rates outputs against constitutional principles → reward model → RL. Scales better but depends on AI judgment quality.

Explanation

Few-shot prompting (Brown et al., 2020): including k demonstrations in the prompt dramatically improves performance on novel tasks without weight updates. The LLM infers the task from examples via in-context learning.

Explanation

Catastrophic forgetting (McCloskey 1989) is the main challenge in continual/lifelong learning. Solutions: EWC (penalize changes to important weights), replay buffers (mix old and new data), modular networks, PackNet (parameter isolation).

Explanation

The derivative of the L2 penalty term (λw²) with respect to w is 2λw — proportional to w, so the shrinkage effect diminishes near zero. The derivative of the L1 penalty (λ|w|) is a constant ±λ, which keeps pushing small weights toward zero until they reach it exactly, producing sparse solutions useful for feature selection.

Explanation

Expected squared error decomposes into bias² + variance + irreducible error (often written as σ², the noise variance in the true relationship between inputs and outputs). Because this noise is intrinsic to the data-generating process, it represents a theoretical floor on performance that no amount of additional data, model capacity, or tuning can remove.