A. BPTT removes recurrence from the model permanently B. BPTT is only used for CNN pooling C. Backpropagation through time unfolds an RNN across time steps to compute gradients D. BPTT is a database backup method
Correct Answer: C. Backpropagation through time unfolds an RNN across time steps to compute gradients
A. Benchmarks replace training algorithms B. Benchmarks remove the need for test data C. Standard benchmarks help compare architectures on common tasks and datasets D. Benchmarks are private weights inside a model
Correct Answer: C. Standard benchmarks help compare architectures on common tasks and datasets
A. DropConnect always drops the whole dataset B. DropConnect is the same as max pooling C. DropConnect is used only for final accuracy reporting D. DropConnect randomly drops weights rather than neuron outputs during training
Correct Answer: D. DropConnect randomly drops weights rather than neuron outputs during training
A. Underfitting means perfect training accuracy B. Underfitting is solved by removing all features C. Underfitting is the same as data leakage D. Underfitting occurs when a model is too simple to capture important patterns
Correct Answer: D. Underfitting occurs when a model is too simple to capture important patterns
A. LSTM networks cannot process sequences B. LSTM networks have no memory cell C. LSTM gates are used only to encrypt data D. LSTM networks use gates to manage long-term dependencies and reduce vanishing-gradient effects
Correct Answer: D. LSTM networks use gates to manage long-term dependencies and reduce vanishing-gradient effects
A. A loss function stores images in GPU memory B. A loss function is the same as a hidden layer C. A loss function cannot guide optimization D. A loss function measures the mismatch between predictions and targets
Correct Answer: D. A loss function measures the mismatch between predictions and targets
A. ResNet uses skip connections to help train very deep networks B. ResNet removes all nonlinear activations C. ResNet is a recurrent architecture only D. ResNet requires no parameters
Correct Answer: A. ResNet uses skip connections to help train very deep networks
A. A multilayer perceptron uses layers of neurons with nonlinear activations B. An MLP is a database indexing method C. An MLP has no trainable weights D. An MLP can only process images
Correct Answer: A. A multilayer perceptron uses layers of neurons with nonlinear activations
A. GPUs accelerate deep learning by performing many parallel numerical operations B. GPUs make datasets unnecessary C. GPU programming replaces all model evaluation D. GPUs are used only for printing predictions
Correct Answer: A. GPUs accelerate deep learning by performing many parallel numerical operations
A. Deep learning learns layered representations from raw data B. It relies only on one manually written rule C. It can never use gradient descent D. It only works with tabular data
Correct Answer: A. Deep learning learns layered representations from raw data
A. GoogleNet is a database recovery method B. GoogleNet/Inception uses parallel filter operations to capture features at multiple scales C. GoogleNet contains only one fully connected layer and no convolutions D. Inception modules forbid 1x1 convolutions
Correct Answer: B. GoogleNet/Inception uses parallel filter operations to capture features at multiple scales
A. Activation functions are used only in the output layer B. Activation functions introduce nonlinearity into neural networks C. Activation functions store files on disk D. Activation functions always remove gradients
Correct Answer: B. Activation functions introduce nonlinearity into neural networks
