Artificial Intelligence (AI) MCQ Questions and Answers

Artificial Intelligence (AI) MCQ Questions and Answers

91. What is Batch Normalization in neural networks?

  1. A data preprocessing technique that normalizes the entire training dataset before training
  2. A technique that normalizes the inputs to each layer within a mini-batch during training, stabilizing the training process, enabling higher learning rates, and reducing sensitivity to weight initialization
  3. A technique for normalizing the batch size used during neural network training
  4. A method for dividing a large dataset into normalized batches for parallel training

Answer : B
Explanation: Batch Normalization (BN), introduced by Ioffe and Szegedy (2015), normalizes the activations of each layer within a mini-batch to have zero mean and unit variance, then applies learnable scale and shift parameters. Benefits: Accelerates training (allows higher learning rates), reduces sensitivity to weight initialization, acts as a mild regularizer (reducing dropout dependence), and helps mitigate the internal covariate shift problem. BN is now a standard component in most deep CNN architectures (ResNet, VGG with BN, EfficientNet). Layer Normalization is a variant used in Transformers since BN doesn’t work well with variable-length sequences.

92. What is an AI Hallucination?

  1. A visual illusion experienced by users when interacting with AI-generated imagery for extended periods
  2. When an AI model — particularly an LLM — confidently generates information that is factually incorrect, fabricated, or not grounded in reality
  3. A technique where AI models imagine possible future scenarios to improve their predictions
  4. A deep learning phenomenon where neurons generate unexpected activation patterns

Answer : B
Explanation: AI Hallucination refers to the tendency of Large Language Models to generate plausible-sounding but factually incorrect, fabricated, or nonsensical content with apparent confidence. Examples include: inventing non-existent academic papers with fake citations, providing wrong historical facts, generating incorrect medical information, or creating fake legal precedents. Hallucinations occur because LLMs generate text based on statistical patterns — they are optimized to produce fluent, coherent text, not necessarily accurate information. Mitigation strategies include Retrieval Augmented Generation (RAG), better training data, RLHF, and uncertainty quantification.

93. What is Retrieval Augmented Generation (RAG) in AI?

  1. A technique for retrieving previously generated AI outputs from a cache for faster responses
  2. An AI architecture that combines a retrieval system (searching relevant documents from a knowledge base) with a generative model (LLM) to produce grounded, factually accurate responses
  3. A data augmentation technique that retrieves similar training examples to improve model robustness
  4. A method for augmenting LLM training data by retrieving content from the internet in real time

Answer : B
Explanation: RAG (Retrieval Augmented Generation) addresses LLM hallucination and knowledge cutoff limitations by combining retrieval and generation. When a user asks a question: (1) The retrieval component searches a knowledge base (documents, databases) for relevant information using semantic search (often via vector embeddings). (2) Retrieved context is injected into the LLM’s prompt. (3) The LLM generates a response grounded in the retrieved facts. RAG allows LLMs to access up-to-date, domain-specific information without retraining. It is widely used in enterprise AI applications, customer support chatbots, and document Q&A systems.

94. What are Embeddings in AI and NLP?

  1. Physical chips embedded in AI hardware to accelerate neural network computations
  2. Dense numerical vector representations of words, sentences, images, or other data that capture semantic meaning and relationships, enabling AI models to process and compare them mathematically
  3. A technique for embedding watermarks into AI-generated content for copyright protection
  4. The process of embedding neural networks inside other neural networks for ensemble learning

Answer : B
Explanation: Embeddings convert discrete data (words, entities, images) into dense continuous vector representations in a high-dimensional space where semantically similar items are positioned close together. Word embeddings (Word2Vec, GloVe, FastText) capture semantic relationships — “king” – “man” + “woman” ≈ “queen”. Sentence embeddings capture meaning of entire sentences. Image embeddings from CNNs capture visual features. Embeddings are the foundation of modern NLP (BERT, GPT use token embeddings), recommendation systems (user and item embeddings), and vector databases used in RAG systems. They allow AI to perform semantic search, clustering, and similarity comparison.

95. What is the difference between AI and Automation?

  1. Automation uses AI; AI cannot be used for automation tasks
  2. Automation follows predefined rules to perform repetitive tasks without human intervention; AI learns from data and can handle new situations, make decisions, and improve over time — AI can power intelligent automation but is broader than traditional automation
  3. AI and Automation are identical — the terms are completely interchangeable
  4. Automation is more advanced than AI and is a superset of all AI capabilities

Answer : B
Explanation: Traditional Automation executes predefined, rule-based tasks with no learning capability — like a factory robot that welds the same spot repeatedly or a scheduled script that runs reports. AI goes further by learning from data, adapting to new situations, and making decisions in ambiguous scenarios. Intelligent Automation (IA) combines traditional automation with AI — for example, RPA (Robotic Process Automation) enhanced with NLP to process unstructured documents. AI-powered automation handles tasks that require judgment, like customer service (chatbots), document processing (intelligent document recognition), and quality inspection (computer vision defect detection).

96. What is Federated Learning in AI?

  1. A machine learning approach where all training data is federated to a central server for processing
  2. A privacy-preserving machine learning approach where models are trained across multiple decentralized devices or servers without sharing raw data, keeping data local while sharing only model updates
  3. A distributed computing method for federating large AI models across multiple data centers
  4. A learning technique where multiple AI models vote to produce a federated final prediction

Answer : B
Explanation: Federated Learning, introduced by Google in 2017, trains AI models across many devices (smartphones, hospitals, banks) without transferring sensitive raw data to a central server. Each device trains a local model on its data, then shares only model updates (gradients) — not the data itself — with a central server that aggregates them into a global model. This preserves data privacy, reduces communication costs, and complies with regulations like GDPR. Applications include Google’s mobile keyboard predictions (Gboard), healthcare AI across hospitals that cannot share patient records, and fraud detection across banks.

97. What is an AI Agent in modern agentic AI systems?

  1. A human employee who acts as an interface between a company and its AI vendor
  2. An AI system that can autonomously plan, make decisions, use tools, and execute multi-step tasks to achieve a goal with minimal human intervention
  3. A software module within a neural network that processes individual data samples
  4. A customer service representative trained using AI-generated scripts and responses

Answer : B
Explanation: Modern AI Agents (also called Agentic AI) are LLM-powered systems that can autonomously complete complex, multi-step tasks. They are characterized by: Goal-directed behavior (given a high-level objective), Tool use (calling web search, code execution, APIs, databases), Memory (retaining context across steps), Planning (breaking goals into sub-tasks), and Self-correction (re-evaluating and adjusting when steps fail). Examples include AutoGPT, Claude Computer Use, and OpenAI’s Operator. Multi-agent systems involve multiple AI agents collaborating on complex tasks. Agentic AI represents the current frontier of AI application development in 2025-2026.

98. What is the concept of AI Safety?

  1. The physical safety regulations for installing AI hardware in data centers
  2. A research field focused on ensuring AI systems behave safely, reliably, and in alignment with human values — preventing unintended harmful outcomes as AI becomes more capable
  3. Cybersecurity measures to protect AI systems from hacking and data breaches
  4. Quality assurance testing procedures applied to AI software before product release

Answer : B
Explanation: AI Safety is a research field addressing risks from increasingly capable AI systems. Key concerns include: Alignment Problem (ensuring AI goals align with human values), Robustness (AI behaving reliably under distributional shift and adversarial inputs), Interpretability (understanding why AI makes certain decisions), Scalable Oversight (maintaining human control as AI becomes more capable), and Catastrophic Risk (preventing AI from taking harmful large-scale actions). Organizations like Anthropic, DeepMind Safety, OpenAI Safety team, and MIRI focus on AI safety research. Constitutional AI and RLHF are current safety techniques used by leading AI labs.

99. What is Multi-Agent System (MAS) in AI?

  1. A system where multiple human agents work alongside a single AI model
  2. A system composed of multiple autonomous AI agents that interact with each other and their environment to collectively solve problems that are too complex for a single agent
  3. A neural network architecture that uses multiple attention heads simultaneously
  4. A hardware system with multiple AI processing chips for parallel computation

Answer : B
Explanation: A Multi-Agent System (MAS) consists of multiple autonomous agents that perceive their environment, make decisions, and interact with other agents to achieve individual or collective goals. Agents can cooperate (working toward shared goals), compete (each pursuing its own goals), or both. MAS enables solving complex distributed problems through agent specialization and parallel processing. Applications include: traffic management systems, supply chain optimization, financial trading platforms, online auctions, robotic swarms, and game AI. Modern LLM-based multi-agent frameworks (like AutoGen and CrewAI) allow multiple AI agents to collaborate on software development, research, and content creation tasks.

100. What is the difference between Supervised Learning and Self-Supervised Learning?

  1. Supervised learning requires no human involvement; self-supervised learning requires constant human supervision
  2. Supervised learning uses human-labeled data; self-supervised learning automatically generates its own supervision signal from unlabeled data by creating a pretext task, enabling training on massive datasets without expensive manual labeling
  3. Self-supervised learning is a subset of supervised learning using automated labeling tools
  4. Both approaches require identical amounts of labeled training data to achieve similar performance

Answer : B
Explanation: Supervised Learning requires human-labeled data — expensive and time-consuming to create at scale. Self-Supervised Learning generates supervision signals automatically from unlabeled data through pretext tasks. Examples: BERT predicts masked words in a sentence (Masked Language Modeling) without human labels. GPT predicts the next token. SimCLR/MoCo learn image representations by treating different augmentations of the same image as positives. Self-supervised learning has enabled training on billions of unlabeled examples, producing powerful foundation models (LLMs, CLIP, DALL-E) that can be fine-tuned for specific tasks with minimal labeled data. It is arguably the most impactful technique driving modern AI progress.

Pages: 1 2 3 4 5 6 7 8 9 10