sem6_ai_5

Which network shares parameters locally?

• Fully connected dense model

• Recurrent sequential memory network

• Generative style adversarial network

• Convolutional weight sharing network

What main challenge motivates CNNs?

• Linear decision boundary limitation

• Excessive handcrafted feature extraction

• Low dimensional image vectors

• Spatial information preservation need

Which layer extracts edge features?

• Fully flattened pixel vector

• Softmax categorical probability layer

• Average pooling subsample layer

• Convolution sliding kernel layer

Stride primarily controls what?

• Kernel element numerical size

• Activation non-linearity performance

• Pooling window average computation

• Filter horizontal vertical shift

Zero padding mainly prevents what?

• Gradient exploding during training

• Parameter count uncontrolled growth

• Vanishing activations in sigmoid

• Border information progressive loss

Which pooling keeps strongest signal?

• Global mean intensity aggregation

• Random fractional downsampling method

• Average neighbourhood value summarisation

• Max neighbourhood value selection

Pooling typically halves which dimension?

• Channel feature map depth

• Epoch training iteration count

• Kernel parameter storage size

• Spatial width and height

ReLU advantages include which point?

• Outputs zero-centred activations

• Always avoids dying problem

• Requires expensive exponential computation

• Mitigates vanishing gradient issue

Leaky ReLU addresses what drawback?

• Exploding gradient unstable updates

• Excessively sparse positive activations

• Probability distribution misalignment error

• Dead inactive neuron phenomenon

Sigmoid is discouraged inside CNNs because?

• Produces sparsity during inference

• Requires negative input normalisation

• Forces zero padding everywhere

• Suffers severe gradient vanishing