A. CuDNN is a replacement for all programming languages B. CuDNN provides optimized GPU routines for deep neural network operations C. CuDNN is a type of loss function D. CuDNN is a data labeling tool only
Correct Answer: B. CuDNN provides optimized GPU routines for deep neural network operations
A. Shallow learning cannot perform classification B. Deep models usually contain multiple hidden layers C. Shallow models always contain more layers than deep models D. Deep learning removes the need for training data
Correct Answer: B. Deep models usually contain multiple hidden layers
A. Computer vision requires no evaluation metrics B. CNNs are unrelated to computer vision C. Deep learning can be applied to image classification, detection, and segmentation D. Deep learning cannot process images
Correct Answer: C. Deep learning can be applied to image classification, detection, and segmentation
A. Backpropagation requires no loss function B. Backpropagation is only a data augmentation method C. Backpropagation computes gradients by applying the chain rule through the network D. Backpropagation stores images in convolution kernels
Correct Answer: C. Backpropagation computes gradients by applying the chain rule through the network
A. A GAN is used only for database indexing B. A GAN cannot generate samples C. A GAN trains a generator and discriminator in an adversarial setting D. A GAN uses only one network with no loss
Correct Answer: C. A GAN trains a generator and discriminator in an adversarial setting
A. A training set contains no labels in supervised learning B. A training set is the final exam dataset C. A training set is used to fit model parameters D. A training set is used only after deployment
Correct Answer: C. A training set is used to fit model parameters
A. Speech recognition is impossible with neural networks B. Speech recognition uses only database joins C. Deep models cannot process audio signals D. Deep learning can map acoustic or sequential features to speech units or text
Correct Answer: D. Deep learning can map acoustic or sequential features to speech units or text
A. Gradient descent updates weights randomly without gradients B. Gradient descent is a file compression algorithm C. Gradient descent only applies to databases D. Gradient descent updates parameters in the direction that reduces loss
Correct Answer: D. Gradient descent updates parameters in the direction that reduces loss
A. Data augmentation means deleting all training images B. Data augmentation is used only after deployment C. Data augmentation makes validation unnecessary D. Data augmentation creates modified training examples to improve robustness
Correct Answer: D. Data augmentation creates modified training examples to improve robustness
A. A test set is used to update weights every epoch B. A test set should be identical to the training set C. A test set is used for data normalization only D. A test set estimates performance on unseen data
Correct Answer: D. A test set estimates performance on unseen data
A. Deep learning is used for language modeling, translation, and text classification B. NLP models cannot use embeddings C. Text classification is not a machine learning task D. NLP cannot involve sequence models
Correct Answer: A. Deep learning is used for language modeling, translation, and text classification
A. The learning rate controls the step size of parameter updates B. The learning rate is the number of output classes C. The learning rate is the dataset size D. The learning rate disables optimization
Correct Answer: A. The learning rate controls the step size of parameter updates