Neural Network Deep Learning Architectures 3

Question 1

What is the key innovation of DenseNet compared to ResNet?

Accepted Answer

Connecting every layer to every subsequent layer

Answer

Using wider layers instead of deeper ones

Answer

Removing all pooling layers

Answer

Using separable convolutions throughout

Question 2

In the context of deep learning, what is a hypernetwork?

Accepted Answer

A network whose weights are generated by another network

Answer

A network trained on hyperparameter tuning

Answer

A network with more than 100 layers

Answer

A network using hyperbolic activation functions

Question 3

What does 'weight sharing' mean in the context of convolutional neural networks?

Accepted Answer

The same filter weights are applied across all spatial positions

Answer

Multiple GPUs share the same gradients

Answer

Weights are shared between encoder and decoder

Answer

Two networks share a common layer

Question 4

Which architecture is most commonly used as the backbone for object detection models like YOLO and Faster R-CNN?

Accepted Answer

Convolutional neural network

Answer

Recurrent neural network

Answer

Transformer encoder only

Answer

Fully connected network

Question 5

What is the purpose of the projection head in contrastive self-supervised learning frameworks like SimCLR?

Accepted Answer

To map representations to a space where contrastive loss is applied

Answer

To classify images into categories

Answer

To generate augmented views of the data

Answer

To apply attention across batch samples

Question 6

What does 'model distillation' (knowledge distillation) transfer from teacher to student network?

Accepted Answer

Soft probability distributions (soft labels) over classes

Answer

The teacher's architecture

Answer

The teacher's optimizer state

Answer

The teacher's training data