Revision

In gradient descent for linear regression, the learning rate affects ​ (a)  

• convergence speed and model accuracy
• model accuracy
• convergence speed

Arrange the steps of the backpropagation algorithm used to train a neural network in the correct order.

• 1. Forward pass

• 1. Compute loss

• 1. Update weights

• 1. Calculate gradients

Match each regularization technique used in training neural networks with its primary purpose or effect on the model.

• Randomly deactivates neurons

• L1 Regularization

• Normalizes layer inputs

• Dropout

• Punishes large weights

• Batch Normalization

The vanishing gradients problem in deep neural networks is most mitigated by:

• Sigmoid activation

• ReLU + Batch Normalization

• Increasing learning rate

• L2 regularization

Arrange the components of a Convolutional Neural Network (CNN) in the correct order of data flow during forward propagation.

• Softmax Activation

• Pooling Layer

• Fully Connected Layer

• Convolutional Layer

What is the primary purpose of subsampling (also known as pooling) in a Convolutional Neural Network?

• Reduce the spatial dimensions of feature maps

• Prevent overfitting by randomly disabling neurons

• Initialize the network’s weights

• Increase the resolution

• Improve the accuracy of predictions directly

Given a 7×7 input image, a 3×3 kernel, a stride of 1, and no padding (padding = 0), what is the size of the output feature map after applying the convolution? Use the formula (input_width - kernel_size + 2*padding)/stride + 1.

The Bag-of-Words text representation method ignores:

• Word order

• Vocabulary size

• Word frequency

• Syntax and Grammar

Arrange the following operations in the correct chronological order that happen during a single LSTM timestep.

• 1. Candidate cell state is generated.

• Output gate computes the hidden state

• 1. Input gate filters new candidate values.

• 1. Cell state is updated.

• 1. Forget gate decides what to remove from the cell state.

How does the attention mechanism in Transformers improve upon traditional RNNs?

• enables parallel processing

• eliminates the need for positional encodings

• uses fewer parameters

• models long-range dependencies