Artificial Intelligence Development Flashcards

Question 1

What does the term artificial intelligence generally refer to?

Accepted Answer

Artificial intelligence generally refers to computer systems designed to perform tasks that normally require human intelligence, such as understanding language, reasoning, learning from experience, and making decisions. The field is broad, and it includes many different methods rather than a single technique. Success depends on how well systems complete these tasks.

Question 2

What is the difference between narrow AI and general AI?

Accepted Answer

Narrow AI refers to systems built for a specific task, such as translating text or recommending videos. General AI, sometimes called strong AI, would perform a wide range of intellectual tasks with human-like flexibility. Today's deployed systems are narrow, while general AI remains a theoretical goal that researchers debate.

Question 3

What is the role of an algorithm in an AI system?

Accepted Answer

An algorithm is a step-by-step procedure that transforms inputs into outputs. In AI systems, algorithms define how a program learns patterns, makes decisions, or searches for solutions. Developers choose or design algorithms based on the problem, the available data, and the computing resources. The same goal can often be reached by several different algorithms.

Question 4

What is symbolic AI and how does it represent knowledge?

Accepted Answer

Symbolic AI represents knowledge with symbols and rules that a program manipulates through logical inference. A developer might encode facts such as every bird has feathers, then write rules that derive new conclusions from them. This approach shaped early research and remains useful for planning, formal reasoning, and systems where explanations must be transparent.

Question 5

What is the Turing test, and what does it try to measure?

Accepted Answer

The Turing test is a proposed evaluation in which a judge converses by typed messages with a hidden machine and a hidden human. If the judge cannot reliably tell which is the machine, the machine is said to pass. Critics note that the test measures imitation of human conversation, not understanding.

Question 6

What is the difference between AI, machine learning, and deep learning?

Accepted Answer

Artificial intelligence is the broadest term, covering any technique that makes machines behave intelligently. Machine learning is a subset of AI in which systems improve performance by learning patterns from data rather than following only hand-written rules. Deep learning is a subset of machine learning that uses neural networks with many layers to learn representations.

Question 7

What does it mean for an AI agent to act rationally?

Accepted Answer

An agent acts rationally when it selects the action expected to best achieve its goals, given what it has perceived so far. An agent is any system that senses its environment and acts on it, such as a thermostat or a chess program. Rationality is judged by outcomes relative to goals, not by human-like thinking.

Question 8

What is a heuristic, and why do AI systems use one in search?

Accepted Answer

A heuristic is a rule of thumb that estimates how promising a choice is, allowing a system to search large possibilities without guaranteeing the best answer. A route planner, for instance, may estimate distance to the goal to choose which path to explore. Heuristics trade some accuracy for speed.

Question 9

What is supervised learning and what kind of data does it need?

Accepted Answer

Supervised learning trains a model on labeled examples, where each input is paired with the correct output. The model adjusts its internal parameters to reduce the gap between its predictions and those known answers. Spam filtering and house price prediction are common examples of this approach.

Question 10

What is unsupervised learning and what is it typically used for?

Accepted Answer

Unsupervised learning finds structure in data that has no labeled answers. Algorithms look for groups, patterns, or compact representations among the examples themselves. Common uses include customer segmentation, which groups shoppers by similar behavior, and dimensionality reduction, which summarizes many features into fewer ones.

Question 11

What is the difference between training data and test data?

Accepted Answer

Training data is the portion of a dataset used to fit a model's parameters. Test data is a separate portion held back until training is complete, and it is used to estimate how well the model performs on examples it has never seen. Keeping these sets apart prevents overly optimistic results.

Question 12

What does overfitting mean, and why is it a problem?

Accepted Answer

Overfitting occurs when a model learns the noise and quirks of its training data so closely that it performs poorly on new data. An overfit model looks excellent on the examples it saw but fails to generalize. Techniques such as collecting more data, simplifying the model, and using regularization help reduce overfitting.

Question 13

What is a loss function and what role does it play in training?

Accepted Answer

A loss function is a formula that measures how far a model's predictions are from the desired outputs, producing a single number that represents error. During training, the optimization process adjusts parameters to make this number smaller. Choosing an appropriate loss function depends on the task, such as regression or classification.

Question 14

How does gradient descent help a model improve during training?

Accepted Answer

Gradient descent is an optimization method that repeatedly nudges model parameters in the direction that most reduces the loss. The gradient indicates how the loss changes as each parameter changes, so the algorithm moves downhill step by step. The size of each step is controlled by a setting called the learning rate.

Question 15

Why do developers use a validation set separate from the test set?

Accepted Answer

A validation set is a portion of data used during development to compare model designs and tune settings such as the learning rate. Because developers choose settings based on validation results, the test set stays untouched and gives a fairer final estimate. Using one set for both tasks would quietly inflate reported performance.

Question 16

What are hyperparameters and how do they differ from parameters?

Accepted Answer

Hyperparameters are settings chosen by the developer before training begins, such as the learning rate or the number of layers. Parameters, by contrast, are values the model learns from data during training, such as the weights in a neural network. Hyperparameters are typically tuned by trying several values and comparing validation results.

Question 17

What is an artificial neuron in a neural network?

Accepted Answer

An artificial neuron is a simple mathematical unit that receives numerical inputs, multiplies each by a weight, adds a bias, and passes the total through an activation function. The output becomes an input for neurons in the next layer. Although loosely inspired by biology, they are far simpler than real neurons.

Question 18

What is a neural network layer, and what are hidden layers?

Accepted Answer

A layer is a group of neurons that process inputs together and pass results forward. The input layer receives raw features, the output layer produces the final prediction, and any layers between them are called hidden layers. Hidden layers allow the network to build intermediate representations that combine simpler features into more abstract ones.

Question 19

What is an activation function and why is it needed?

Accepted Answer

An activation function is a nonlinear function applied to a neuron's weighted sum to produce its output. Without nonlinearity, stacking many layers would still compute only a linear function, so the network could not model complex relationships. Common examples include the sigmoid, tanh, and ReLU functions.

Question 20

How does backpropagation help a neural network learn from its errors?

Accepted Answer

Backpropagation is an algorithm that computes how much each weight contributed to the error, using the chain rule of calculus. After a forward pass produces a prediction and a loss, the error is propagated backward through the layers. Each weight is then adjusted, usually by gradient descent, to reduce that error on future passes.

Question 21

What is a convolutional neural network typically used for?

Accepted Answer

A convolutional neural network is designed to process grid-like data, such as images, by sliding learned filters across the input to detect local patterns. Early layers tend to respond to edges and textures, while deeper layers combine these into shapes and objects. They support tasks such as image classification and object detection.

Question 22

What kind of data are recurrent neural networks designed to handle?

Accepted Answer

Recurrent neural networks are designed for sequential data, where the order of inputs matters, such as text, speech, or time series. They keep a hidden state that carries information from earlier steps forward, so each output can depend on what came before. Transformers now handle many language tasks more effectively.

Question 23

What is transfer learning and why is it useful?

Accepted Answer

Transfer learning reuses a model trained on one task as the starting point for a related task. Knowledge such as general visual features or language patterns is retained, so the new model needs less labeled data and less training time. Developers often fine-tune only the final layers on their own data.

Question 24

What makes training deep neural networks so computationally demanding?

Accepted Answer

Training deep networks involves repeatedly multiplying large matrices of numbers across millions or even billions of parameters, for many passes over the data. Specialized processors such as GPUs handle these parallel calculations far faster than standard processors. The cost rises with model size, dataset size, and the number of training iterations.

Question 25

Why is data quality so important in AI development?

Accepted Answer

Data quality strongly influences what a model can learn, because a model reflects the patterns, errors, and gaps in its training examples. Incomplete, mislabeled, or noisy data can produce unreliable predictions even when the algorithm is well designed. Careful collection, cleaning, and documentation are therefore central parts of development work.

Question 26

What is data labeling, and why does it matter?

Accepted Answer

Data labeling is the process of attaching correct answers or descriptive tags to raw examples, such as marking which photos contain a bicycle. Supervised learning depends on these labels, so label accuracy directly limits model accuracy. Labeling is often slow and costly, and annotators may disagree, which means clear written guidelines are essential.

Question 27

What is class imbalance, and why can it mislead a model's accuracy?

Accepted Answer

Class imbalance occurs when some categories appear much more often than others, such as fraudulent transactions making up a tiny share of all transactions. A model can achieve high accuracy by always predicting the common class, while failing on the rare cases that matter most. Precision and recall give a clearer picture.

Question 28

What does a confusion matrix show about a classifier's results?

Accepted Answer

A confusion matrix is a table that counts how often a classifier's predictions match or miss each true class. Its layout lets you read counts of true positives, false positives, true negatives, and false negatives. These counts are the basis for metrics such as precision, recall, and accuracy.

Question 29

What do precision and recall each measure in a classifier?

Accepted Answer

Precision measures the share of positive predictions that are actually correct, answering how trustworthy a positive result is. Recall measures the share of all true positive cases that the model successfully finds, answering how many real cases were caught. The two often trade off against each other.

Question 30

What is k-fold cross-validation and why is it useful with limited data?

Accepted Answer

K-fold cross-validation splits the data into k equal parts, then trains k times, each time holding out a different part for testing and training on the rest. Averaging the results gives a steadier performance estimate. This helps when data is limited, because every example is eventually used for both training and evaluation.

Question 31

What is underfitting, and how does it differ from overfitting?

Accepted Answer

Underfitting occurs when a model is too simple or undertrained to capture the real pattern in the data, so it performs badly on both training and new examples. Overfitting is the opposite problem, where the model fits training data too closely and fails to generalize. Balancing these is a central goal of model selection.

Question 32

What is a baseline model, and why compare against one?

Accepted Answer

A baseline model is a simple reference point, such as always predicting the most common class or using a basic linear rule. Comparing a complex model against a baseline shows whether the added complexity actually improves results. A model that barely beats a trivial baseline may not justify its extra cost.

Question 33

What does deploying an AI model involve beyond training it?

Accepted Answer

Deployment means making a trained model available to users or other software, often through an application programming interface or an embedded product feature. It also requires infrastructure for serving predictions, monitoring performance over time, and updating the model when conditions change. Deployment and maintenance often receive too little planning, which causes problems later.

Question 34

What is model drift, and why does it matter after deployment?

Accepted Answer

Model drift occurs when the real-world data a model sees after deployment changes from the data it was trained on, so its predictions gradually become less accurate. Shifts in customer behavior, language, or sensor conditions are common causes. Monitoring incoming data and retraining the model are common ways to detect and respond to drift.

Question 35

What is algorithmic bias, and how can it arise?

Accepted Answer

Algorithmic bias refers to systematic errors in an AI system that produce unfair or uneven outcomes for certain groups of people. It can arise from unrepresentative training data, from labels that reflect past discrimination, or from design choices that favor some objectives over others. Teams should audit data and results throughout development.

Question 36

Why is explainability important for certain high-stakes AI applications?

Accepted Answer

Explainability is the ability to describe why a model produced a particular output in terms people can understand. It matters in high-stakes settings such as lending, medical diagnosis, or legal decisions, where affected people reasonably want to know the reasons. Explanations also help developers find errors and spot hidden biases.

Question 37

What is privacy risk in AI development, and what is one common safeguard?

Accepted Answer

Privacy risk arises when AI systems train on or reveal personal information, such as health records or private messages, without adequate protection. A common safeguard is data minimization, which means collecting and retaining only the information truly needed for the task. Techniques such as anonymization and access controls also reduce exposure.

Question 38

How do AI developers usually decide whether a model is ready to release?

Accepted Answer

Developers usually decide readiness by checking performance on held-out test data, comparing results against a baseline, and reviewing errors across relevant user groups. They also run safety and fairness evaluations, test the system in realistic conditions, and confirm that the team can monitor it after launch. Release is a judgment informed by these results.

Question 39

What is a human-in-the-loop design, and when is it useful?

Accepted Answer

Human-in-the-loop design places a person in the decision process so people review, correct, or approve AI outputs before they take effect. This approach is useful when errors are costly or when judgment, accountability, or context matters, such as medical triage support or content moderation. It also produces corrected examples for future training.

Question 40

What is the purpose of documenting an AI system, such as with a model card?

Accepted Answer

Documentation records what a system is for, what data trained it, how it was evaluated, and where it is known to perform poorly. A model card is one common format for this information, written so that users, auditors, and future developers can judge whether the model fits their needs. It supports accountability.

Artificial Intelligence Development Flashcards

1Foundations of Artificial Intelligence

2Machine Learning Fundamentals

3Neural Networks and Deep Learning

4Data and Model Evaluation

5Applied AI and Responsible Use

Learning Resources