Artificial Intelligence(AI) Interview Questions 2025

This is a fundamental challenge in supervised learning, describing the conflict between two types of errors a model can make:

• Bias is the error from overly simplistic assumptions. A high-bias model is too simple and fails to capture the underlying patterns in the data, leading to underfitting.
• Variance is the error from too much complexity. A high-variance model is too sensitive to the noise in the training data, leading to overfitting.

The tradeoff means that as you decrease a model's bias (by making it more complex), you typically increase its variance, and vice-versa. The goal is to find a balance that minimizes the total error on unseen data.

Regularization is a set of techniques used to prevent overfitting in machine learning models, especially in Logistic Regression and neural networks. It works by adding a penalty term to the model's loss function, which discourages the model from assigning excessive weights to features.

• L1 Regularization (Lasso): Adds a penalty equal to the absolute value of the weights. It can shrink some weights to exactly zero, effectively performing feature selection.
• L2 Regularization (Ridge): Adds a penalty equal to the square of the weights. It forces weights to be small but rarely shrinks them to zero.

Gradient Descent is an optimization algorithm used to find the minimum of a function, which in machine learning is the loss function. The main idea is to take repeated steps in the opposite direction of the gradient (or slope) of the function at the current point, as this is the direction of steepest descent.

An activation function is a mathematical function applied to the output of a neuron. Its primary purpose is to introduce non-linearity into the network. Without non-linear activation functions, a deep neural network would behave just like a single-layer linear model, unable to learn complex patterns found in data like images or speech.

Common examples include Sigmoid, Tanh, and ReLU (Rectified Linear Unit), with ReLU being the most popular choice in modern networks due to its efficiency.

A loss function (or cost function) quantifies how "wrong" a model's prediction is compared to the actual label. It calculates a single number representing the error for the current state of the model. The entire goal of the training process is to adjust the model's parameters (weights) to minimize this loss function's value. Different tasks use different loss functions (e.g., Mean Squared Error for regression, Cross-Entropy for classification).

These are evaluation metrics used for classification tasks, especially when dealing with imbalanced classes:

• Precision: Of all the positive predictions the model made, how many were actually correct? (Focuses on minimizing False Positives).
• Recall (Sensitivity): Of all the actual positive instances, how many did the model correctly identify? (Focuses on minimizing False Negatives).
• F1-Score: The harmonic mean of Precision and Recall, providing a single score that balances both metrics.

Example: In cancer detection, Recall is critical because you want to find all actual cancer cases (minimizing missed diagnoses/False Negatives).

K-Fold Cross-Validation is a resampling procedure used to get a more reliable estimate of a model's performance on unseen data. It helps ensure the model is robust and that its performance isn't just due to a lucky split of the training and test data.

Parameters are internal variables that the model learns on its own from the training data. Their values are the output of the training process. Example: The weights and biases in a neural network.

Hyperparameters are external, high-level settings that are configured by the data scientist before the training process begins. They control how the model learns. Example: The learning rate, the number of layers in a neural network, the 'K' in K-Fold Cross-Validation.

Ensemble methods are techniques that combine the predictions from multiple machine learning models to produce a more accurate and robust prediction than any single model. The idea is that "many heads are better than one."

• Bagging (Bootstrap Aggregating): Trains multiple models in parallel on different random subsets of the data. Example: Random Forest, which builds many Decision Trees.
• Boosting: Trains multiple models sequentially, where each new model tries to correct the errors made by the previous ones. Example: AdaBoost, Gradient Boosting Machines (GBM).

Backpropagation (short for "backward propagation of errors") is the core algorithm for training artificial neural networks. After the network makes a prediction (a "forward pass"), backpropagation calculates the gradient of the loss function with respect to the network's weights. It does this by propagating the error backward from the output layer to the input layer. This gradient is then used by the optimization algorithm (like Gradient Descent) to update the weights in a way that minimizes the error.

These are significant challenges that arise during the training of deep artificial neural networks through backpropagation.

• Vanishing Gradients: Occurs when gradients become extremely small as they are propagated backward through the layers. This makes the weights in the earlier layers update very slowly, or not at all, effectively stopping the network from learning.
• Exploding Gradients: The opposite scenario, where gradients become excessively large, leading to huge weight updates and causing the training process to become unstable and diverge.

Solutions: Using activation functions like ReLU, implementing residual connections (ResNets), and using gradient clipping are common strategies to combat these issues.

Transfer Learning is a technique where a model developed for one task is reused as the starting point for a model on a second, related task. Instead of training a new model from scratch, you use the "knowledge" (weights and features) learned from a pre-trained model.

Use Cases: It's extremely useful when your target task has a limited amount of data. For example, you can use a powerful model pre-trained on the huge ImageNet dataset (millions of images) and then fine-tune its final layers for your specific image classification task, like identifying different types of flowers, which might only have a few thousand images. It's a core concept in many AI courses.

An imbalanced dataset (e.g., 99% non-fraud vs. 1% fraud transactions) can cause a model to be biased towards the majority class. Several techniques can be used:

• Use Appropriate Metrics: Don't use accuracy. Use metrics like Precision, Recall, F1-Score, or AUC-ROC that provide a better picture of performance.
• Resampling Techniques: Modify the dataset by either oversampling the minority class (e.g., using SMOTE) or undersampling the majority class.
• Generate Synthetic Data: Create new, synthetic examples of the minority class. This is central to many Generative AI Courses.
• Use Different Algorithms: Tree-based algorithms like Random Forest and Gradient Boosting often perform better on imbalanced data.

Both are classes of statistical models, but they learn different things from the data.

• A Discriminative Model learns the decision boundary between different classes. It models the conditional probability, P(Y|X). It's good for classification tasks. Example: Support Vector Machines (SVM), Logistic Regression.
• A Generative Model learns the actual distribution of each class. It models the joint probability, P(X, Y), and can be used to generate new data samples. Example: Naive Bayes, Generative Adversarial Networks (GANs).

The Attention Mechanism is a technique that allows a model to focus on the most relevant parts of the input sequence when producing a specific part of the output sequence. Instead of compressing an entire input sequence into a single fixed-length vector (which can be a bottleneck), attention allows the model to "look back" at the input sequence and assign different "attention scores" or weights to different input words.

Batch size is a crucial hyperparameter that determines the number of samples processed before the model's internal parameters are updated. Using a large batch size has distinct advantages and disadvantages.

• Pros:
  • Stable Gradient Estimate: Larger batches provide a more accurate estimate of the gradient, leading to a smoother and more stable convergence path.
  • Computational Efficiency: Modern hardware (GPUs/TPUs) is optimized for parallel computations, making processing one large batch faster than many small ones.
• Cons:
  • Higher Memory Requirement: All samples in the batch must be loaded into memory, which can be a significant limitation. This requires careful memory management, similar to Garbage Collection in Java.
  • Poorer Generalization: Research suggests large batches can converge to sharp, less robust minima in the loss landscape, while smaller batches find flatter minima that generalize better to new data.

Dimensionality reduction is the process of reducing the number of features (dimensions) in a dataset. It's important for several reasons, chief among them combating the "Curse of Dimensionality," where data becomes very sparse in high dimensions, making models harder to train and more prone to overfitting. It also reduces computational cost and can help with data visualization.

Principal Component Analysis (PCA) is a popular technique that works by identifying a new set of orthogonal axes, called principal components, that capture the maximum amount of variance in the data. By keeping only the first few principal components, we can reduce the number of dimensions while retaining most of the information and the original data's correlation structure.

Explainable AI (XAI) is an area of AI research and practice that focuses on creating systems whose decisions can be understood by humans. It addresses the "black box" problem of complex models like deep neural networks, where it's often unclear why a specific prediction was made.

It's becoming more important for several reasons:

• Trust and Adoption: Users are more likely to trust and adopt AI systems if they understand how they work, a core tenet similar to the Amazon Leadership Principle of "Earn Trust".
• Debugging and Fairness: XAI helps developers identify and correct hidden biases or flaws in their models.
• Regulatory Compliance: Regulations like GDPR in Europe give users a "right to explanation" for decisions made by automated systems.
• Critical Applications: In fields like healthcare and finance, understanding the 'why' behind a decision is often a legal and ethical necessity.

Concept Drift is a phenomenon where the statistical properties of the target variable change over time. This means the relationship between the input features and the output label, which the model learned during training, is no longer valid in the real world.

Example: A fraud detection model trained on historical data may become ineffective when fraudsters develop entirely new methods of committing fraud. The "concept" of fraud has drifted. This is a major challenge in big data analytics where data is constantly streaming.

Detection: The primary way to detect concept drift is through continuous monitoring of the model's performance on live data. A sudden or gradual degradation of key metrics (like F1-score, precision, or recall) is a strong indicator that the model's learned patterns are becoming outdated and that it may need to be retrained on more recent data.

The primary purpose of a pooling layer in a CNN is to perform downsampling—that is, to progressively reduce the spatial size (width and height) of the feature maps. This serves two main benefits:

• Reduces Computational Cost: By shrinking the feature maps, it decreases the number of parameters and computations in the network. This not only speeds up training but also helps to control overfitting.
• Provides Translational Invariance: Pooling makes the feature detection more robust to variations in the location of the feature in the image. For example, Max Pooling (the most common type) takes the maximum value from a patch of pixels. This means the network detects whether a feature is present within a region, rather than exactly where it is.