Artificial Intelligence Interview Questions and Answers (2025) | JaganInfo

Artificial Intelligence is the simulation of human intelligence in machines programmed to think, learn, and make decisions.

Narrow AI, which is designed for specific tasks, and General AI, which can perform any intellectual task a human can.

Machine Learning is a subset of AI focused on building systems that learn from data to improve performance on tasks.

Supervised learning trains models on labeled data to make predictions or classifications.

Unsupervised learning identifies patterns or groupings in unlabeled data.

A neural network is a computational model inspired by human brain structure, used to recognize patterns and solve AI problems.

Deep learning is a subset of machine learning involving neural networks with many layers that learn hierarchical representations.

Overfitting occurs when a model learns noise and details from training data too well, reducing performance on new data.

Reinforcement learning trains models to make sequences of decisions by rewarding desired actions and penalizing undesired ones.

Applications include speech recognition, autonomous vehicles, image recognition, chatbots, recommendation systems, and robotics.

AI is a broad concept of machines simulating intelligence; machine learning is a subset involving learning from data; deep learning is a subset of ML using deep neural networks.

Common algorithms include linear regression, logistic regression, decision trees, support vector machines, and neural networks.

Decision trees split data based on feature conditions to make predictions; interpretable and used for classification and regression.

Bias is error from assumptions in the model, variance is error from sensitivity to training data; balancing both is key to good generalization.

Combining multiple models to improve predictive performance, e.g., Random Forest, Gradient Boosting.

An optimization algorithm that iteratively adjusts model parameters to minimize the loss function.

Activation functions introduce non-linearity; common examples are ReLU, Sigmoid, and Tanh.

Settings like learning rate, number of layers, and batch size set before training a model that affect performance.

Techniques to reduce overfitting by penalizing complex models, such as L1, L2 regularization, and dropout.

A technique to stop training when validation error increases, preventing overfitting.

CNNs are deep learning models effective for image and spatial data analysis, using convolutional layers to extract features.

RNNs are designed for sequential data and use internal states to capture temporal dependencies.

LSTMs are RNN variants addressing vanishing gradient problems for longer sequence learning.

Using a pre-trained model on a large dataset as the starting point for a related task to improve learning efficiency.

Techniques applied to raw data to clean, normalize, and prepare it for modeling.

Loss functions quantify the difference between predicted and actual outputs to guide model optimization.

Algorithms like SGD, Adam, and RMSprop adjust model parameters to minimize loss during training.

When gradients diminish in deep networks during backpropagation, hindering weight updates.

Techniques include resampling, synthetic data generation, and cost-sensitive learning.

Method to normalize layer inputs during training to stabilize and accelerate learning.

A Transformer uses self-attention mechanisms and feed-forward layers to process sequences in parallel, improving efficiency over RNNs.

Self-attention computes a weighted representation of all elements in the sequence to capture dependencies irrespective of distance.

BERT is a bidirectional Transformer model pretrained via masked language modeling that achieves state-of-the-art NLP performance.

The Generative Pre-trained Transformer is a unidirectional Transformer model optimized for text generation tasks.

Models pretrained on large datasets are fine-tuned on smaller, task-specific datasets to improve performance and reduce training time.

A learning paradigm where agents learn to make decisions by interacting with the environment and receiving rewards.

GANs consist of generator and discriminator models competing to produce realistic fake data indistinguishable from real data.

Attention relates inputs regardless of position; positional encoding injects order information to compensate for attention’s position-agnostic nature.

Techniques to understand and explain how AI models make decisions, important for trust and compliance.

Challenges include computational cost, data requirements, model tuning, and deployment complexities.

Using regularization, dropout, data augmentation, early stopping, and transfer learning.

Using containerization, APIs, model monitoring, versioning, and continuous retraining pipelines.

Capability of models to learn from new data without forgetting previously learned knowledge.

Approaches that provide human-understandable explanations for AI decisions to enhance transparency and trust.

Masks prevent the model from attending to padding tokens or future tokens during training.

Ability of models to perform tasks without explicit training examples through generalization.

Attention mechanisms enable parallel processing and efficient handling of long-range dependencies.

Large LMs are pretrained on vast text corpora and fine-tuned on specific tasks or datasets for specialization.

Inputs crafted with subtle perturbations that cause AI models to make incorrect predictions.

Includes biases in data, privacy violations, autonomy, accountability, and the impact on employment.