Artificial Intelligence (AI) MCQ Questions and Answers

Artificial Intelligence (AI) MCQ Questions and Answers

81. What is Overfitting and Underfitting in AI/ML models?

  1. Overfitting means the model runs too slowly; Underfitting means the model runs too fast
  2. Overfitting occurs when a model learns training data too well including noise, failing on new data; Underfitting occurs when a model is too simple to capture the underlying patterns in the data
  3. Overfitting means the model has too many layers; Underfitting means it has too few neurons
  4. Both overfitting and underfitting refer to the same phenomenon with different severity levels

Answer : B
Explanation: Overfitting: the model memorizes training data — including noise — resulting in high training accuracy but poor generalization to new data (high variance). Solutions: more training data, regularization (L1/L2), dropout, early stopping, cross-validation. Underfitting: the model is too simple to capture the data’s complexity — resulting in poor performance on both training and test data (high bias). Solutions: more complex model, more features, longer training, reduce regularization. The ideal model finds the “sweet spot” between bias and variance — this is called the Bias-Variance Tradeoff, one of the most fundamental concepts in machine learning.

82. What is ChatGPT and what technology powers it?

  1. A search engine developed by OpenAI that indexes web pages using AI
  2. A conversational AI chatbot developed by OpenAI, powered by large language models (GPT series) based on the Transformer architecture, capable of generating human-like text responses
  3. A computer vision system that generates image captions using GPT technology
  4. A coding assistant that only writes Python and JavaScript code for developers

Answer : B
Explanation: ChatGPT is a conversational AI chatbot developed by OpenAI, first released in November 2022. It is built on GPT (Generative Pre-trained Transformer) large language models and fine-tuned using RLHF (Reinforcement Learning from Human Feedback) to follow instructions and produce helpful, harmless, and honest responses. ChatGPT can write essays, generate code, answer questions, summarize documents, translate languages, and more. It became one of the fastest-growing applications in history, reaching 100 million users in just two months. It sparked the current wave of Generative AI development across the industry.

83. What is RLHF (Reinforcement Learning from Human Feedback) in AI?

  1. A technique where AI robots learn from human physical demonstrations in real-world environments
  2. A training technique that uses human preferences and feedback to fine-tune AI models to be more helpful, harmless, and aligned with human values
  3. A reinforcement learning method where human trainers play games against AI agents
  4. A data labeling process where humans manually annotate training datasets for AI models

Answer : B
Explanation: RLHF (Reinforcement Learning from Human Feedback) is a training methodology used to align large language models with human intentions. The process: (1) Fine-tune a base LLM using supervised learning on human-written demonstrations. (2) Collect human preference data — humans rank multiple model outputs from best to worst. (3) Train a Reward Model to predict which outputs humans prefer. (4) Use RL (specifically PPO — Proximal Policy Optimization) to fine-tune the LLM to maximize the reward model’s score. RLHF is used by OpenAI (ChatGPT), Anthropic (Claude), and Google (Gemini) to make their models more helpful and safer.

84. What is the difference between AI and Robotics?

  1. AI and Robotics are identical fields — all robots are AI-powered
  2. AI is the intelligence/software that enables machines to think and learn; Robotics is the engineering of physical machines — they can work independently (a robot without AI uses pre-programmed rules; AI without a robot can exist as pure software)
  3. Robotics is more advanced than AI and contains AI as a subset
  4. AI deals only with industrial manufacturing; Robotics deals only with software development

Answer : B
Explanation: AI and Robotics are related but distinct fields. AI is the branch of computer science focused on creating intelligent behavior in machines — it can exist entirely as software (like ChatGPT). Robotics is the engineering discipline focused on designing, building, and programming physical machines (robots). When combined, AI-powered robots can perceive their environment, reason, and act autonomously. However, many robots operate without AI (simple factory arms following pre-programmed movements), and most AI systems have no physical body (recommendation engines, language models). The intersection — AI robotics — represents increasingly important applications in manufacturing, healthcare, and services.

85. What is the Softmax function in neural networks?

  1. A function that softens the ReLU activation to reduce dying neuron problems
  2. An activation function applied to the output layer of multi-class classification models that converts raw scores (logits) into a probability distribution summing to 1
  3. A normalization function applied to all hidden layers to speed up neural network training
  4. A regularization technique that randomly sets some neurons’ outputs to zero during training

Answer : B
Explanation: Softmax converts a vector of raw scores (logits) into probabilities: σ(zᵢ) = e^zᵢ / Σe^zⱼ. Each output is between 0 and 1, and all outputs sum to exactly 1 — making them interpretable as class probabilities. For example, in a 3-class problem, Softmax might output [0.7, 0.2, 0.1], meaning 70% probability for class 1. Softmax is used in the final output layer for multi-class classification. Cross-entropy loss is used alongside Softmax during training. For binary classification, Sigmoid is preferred over Softmax. Softmax is a fundamental concept tested in all AI and deep learning exams.

86. What is AlphaGo and why is it significant in AI history?

  1. A robot developed by Google to play physical board games against human opponents
  2. An AI program developed by DeepMind that defeated the world champion at the ancient board game Go in 2016, demonstrating AI can master extremely complex games once considered beyond machine capability
  3. A chess-playing AI developed by IBM that first defeated a world chess champion in 1997
  4. An AI algorithm developed by OpenAI that solved the protein folding problem in biology

Answer : B
Explanation: AlphaGo, developed by DeepMind (a Google subsidiary), defeated world Go champion Lee Sedol 4-1 in March 2016 — a historic milestone. Go was considered the most complex board game, with more possible positions than atoms in the observable universe, making brute-force impossible. AlphaGo combined deep neural networks (for position evaluation and move selection) with Monte Carlo Tree Search and reinforcement learning. Its successor AlphaGo Zero learned purely through self-play with no human knowledge and became even stronger. AlphaZero then mastered Chess and Shogi in addition to Go, all from scratch.

87. What is the purpose of Dropout in neural network training?

  1. To remove neurons that produce incorrect outputs during the testing phase
  2. A regularization technique that randomly deactivates a fraction of neurons during each training step to prevent overfitting and force the network to learn more robust features
  3. A data preprocessing method that drops incomplete or missing records from the training set
  4. A learning rate scheduling technique that gradually drops the learning rate during training

Answer : B
Explanation: Dropout is a powerful regularization technique introduced by Srivastava et al. (2014). During training, each neuron is randomly “dropped” (set to zero) with a probability p (typically 0.2-0.5), preventing the network from becoming too dependent on any single neuron. This forces the network to develop multiple redundant representations, improving generalization. At inference (testing) time, all neurons are active but their outputs are scaled by (1-p). Dropout is one of the most widely used regularization techniques in deep learning and is applied in CNNs, RNNs, and Transformer models to prevent overfitting.

88. What is Feature Engineering in AI/ML?

  1. The process of designing new neural network architectures with optimal feature detection layers
  2. The process of using domain knowledge to manually create, select, and transform input variables (features) from raw data to improve machine learning model performance
  3. A deep learning technique where features are automatically discovered from raw data
  4. The engineering work involved in building the physical features of AI hardware chips

Answer : B
Explanation: Feature Engineering is the process of using domain expertise to create informative input variables that help ML models perform better. Techniques include: Feature Creation (creating new features from existing ones — e.g., “age” from “birth date”), Feature Transformation (log transformation, normalization, standardization), Feature Selection (removing irrelevant or redundant features), Feature Encoding (one-hot encoding for categorical variables), and Feature Extraction (PCA, embeddings). Good feature engineering can dramatically outweigh model selection — a simple model with great features often beats a complex model with poor features. Deep learning automates much of this process.

89. What is Gradient Descent in AI/ML model training?

  1. An algorithm that descends through multiple gradient levels in a deep neural network
  2. An optimization algorithm that iteratively adjusts model parameters in the direction opposite to the gradient of the loss function to minimize the loss and improve model accuracy
  3. A visualization technique that shows the gradient of model accuracy during training
  4. A data preprocessing technique for handling steep gradients in skewed datasets

Answer : B
Explanation: Gradient Descent is the core optimization algorithm for training neural networks. It calculates the gradient (slope) of the loss function with respect to each model parameter and updates parameters in the opposite direction (downhill) to reduce loss. The learning rate controls step size. Variants: Batch Gradient Descent (uses entire dataset per update — slow but stable), Stochastic Gradient Descent — SGD (uses one sample per update — fast but noisy), Mini-batch GD (uses small batches — balance of both). Modern optimizers like Adam, RMSprop, and AdaGrad adaptively adjust learning rates per parameter for faster, more stable convergence.

90. What is the difference between Narrow AI, AGI, and ASI?

  1. These are three versions of the same AI system with increasing computational power
  2. Narrow AI performs specific tasks; AGI (Artificial General Intelligence) would match human intelligence across all tasks; ASI (Artificial Superintelligence) would surpass human intelligence in all domains — only Narrow AI exists today
  3. AGI is a subset of Narrow AI; ASI is a subset of AGI in the AI capability hierarchy
  4. All three categories currently exist and are deployed in different commercial applications

Answer : B
Explanation: The three levels of AI capability: Narrow AI (Weak AI) — all currently existing AI systems, designed and optimized for one specific task (ChatGPT for text, AlphaGo for Go, FaceID for face recognition). Artificial General Intelligence (AGI) — hypothetical AI with human-level general cognitive ability, able to learn and perform any intellectual task. Does not yet exist. Artificial Superintelligence (ASI) — hypothetical AI that surpasses human intelligence in every domain. Purely theoretical. The development timeline and feasibility of AGI and ASI are actively debated among AI researchers. This is a top-searched AI exam and interview topic.

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