Test your understanding of generative artificial intelligence principles with these easy questions. This quiz covers the basics of neural networks, generative models, activation functions, AI learning methods, and foundational concepts in generative AI.
Which part of an artificial neural network is responsible for receiving the raw input data?
Explanation: The input layer of a neural network receives the raw data features for further processing. The output layer provides the final predictions, while hidden layers transform and extract features. An activation function adds non-linearity but is not a physical 'layer' itself.
In a standard feedforward neural network, what is the main role of hidden layers?
Explanation: Hidden layers in neural networks extract relevant features from the data and perform intermediate computations. The input layer presents raw data, not features. The output layer displays predictions, but does not compute intermediate patterns. Labels are part of the dataset, not the network architecture.
Why are activation functions used in neural networks, such as ReLU or Sigmoid?
Explanation: Activation functions add non-linearity so networks can learn complex, non-linear relationships. They do not store results, decrease accuracy, or sort data. Without non-linearity, networks behave like simple linear models and cannot solve complex tasks.
Which key difference separates generative models from discriminative models?
Explanation: Generative models learn to produce data resembling the training distribution, while discriminative models focus on classifying or labeling inputs. Discriminative models do not generate new data, so option B is incorrect. Both types of models are trained with data, and discriminative models are especially used for classification.
Which of the following is commonly used as a generative model in AI?
Explanation: A Generative Adversarial Network (GAN) is a popular generative model for producing images, text, or other data. Decision Trees and Support Vector Machines are discriminative models. k-Means is a clustering algorithm, not a generative model.
In a neural network, what do we call the parameters that are adjusted during training to minimize errors?
Explanation: Weights are the parameters updated during training in a neural network to reduce the error between predicted and true outputs. Neurons and layers describe the structural components of the network. Activations are the outputs of neurons after applying activation functions.
Which training algorithm is widely used in neural networks to update the weights based on output error?
Explanation: Backpropagation calculates gradients of errors and updates the weights to reduce future errors. Data augmentation changes training inputs for robustness, principal component analysis is used for dimensionality reduction, and clustering groups similar data but does not update weights for prediction.
What is the main function of the output layer in a neural network?
Explanation: The output layer generates the network's final prediction or classification based on preceding computations. Feature extraction occurs mainly in hidden layers. The output layer does not introduce randomness or receive input data directly.
What is the main goal of using gradient descent during neural network training?
Explanation: Gradient descent updates model weights to reduce prediction errors over the training data. It does not increase dataset size, modify activation functions, or shuffle labels. Its focus is efficient optimization of network parameters.
In supervised learning, what kind of data does a generative model receive during training?
Explanation: Supervised learning involves training with both input data and the correct output labels for each example. Unsupervised learning, in contrast, uses only input data with no labels. Providing just outputs or unlabeled, unstructured data does not define supervised settings.
Which one of the following functions is a common activation function in neural networks?
Explanation: Sigmoid is a widely used activation function that transforms input values into a range between 0 and 1. Linear regression and gradient boosting are types of models, not activation functions. Mean squared error is a loss function, not an activation function.
What is a typical application of a generative model?
Explanation: Generative models are often used to create new data samples, such as images, that resemble the training data. Sorting or calculating averages are general data processing tasks, while encoding is more related to compression, not generation.
What happens in a neural network when it 'learns' from data?
Explanation: Learning in neural networks means adapting internal weights based on errors to improve predictions. Structure changing or data creation are not core learning steps. Repeating previous outputs does not help a network learn from new data.
Why must raw input data often be scaled or normalized before training a neural network?
Explanation: Scaling or normalizing input makes sure that features with larger values do not dominate the learning process. Increasing memory, adding bias, or hiding features are not goals of data normalization before training.
What does a loss function measure in a neural network?
Explanation: A loss function quantifies how well the predictions match the correct outputs, guiding training. Number of layers, data size, and initialization methods are not measured by the loss function.
Why is a diverse training dataset important when building generative AI models?
Explanation: A diverse dataset helps the model learn broader patterns and generalize to data it has not seen. Increased speed and repetitive examples do not improve performance, and errors usually decrease with diverse, meaningful data rather than increase.