What is the main purpose of padding in a convolutional neural network?
To increase the number of parameters
To allow filters to cover edge pixels and keep output size larger
To change the activation function
To speed up training
Cross‑entropy loss is primarily used to measure:
The similarity between two probability distributions (true labels vs. predictions)
The amount of overfitting in a model
The speed of convergence during training
The variance of the input data
In image transfer learning, which layers are usually retrained for a new task?
All layers from input to output
Only the early convolutional layers
Only the final layers that learn high‑level features
Only the pooling layers
What is the main purpose of an activation function in a neural network?
To reduce the learning rate
To convert weights into biases
To introduce non-linearity into the model
To normalize the input data