Computer Vision

Which of the following techniques is used to boost model performance by reusing a pretrained network's base in image classification tasks?

• Data Augmentation

• Feature Extraction

• Transfer Learning

• Binary Crossentropy

Answer explanation

Transfer learning is a technique where a pretrained model, usually on a large and comprehensive dataset like ImageNet, is used as the base for a new model. The pretrained base already has learned feature extraction, which can be utilized in the new task, improving performance and reducing the amount of data needed for training.

In a convolutional neural network (CNN), what is the primary function of the convolutional base?

• To determine the class of the image

• To extend the dataset

• To extract features from the image

• To transform outputs to a probability score

Answer explanation

The convolutional base of a CNN is responsible for extracting features from the input image. It does this through a series of convolutional operations, capturing patterns like edges, textures, and shapes, which are essential for the subsequent classification task performed by the dense head of the network.

What is the primary purpose of the ReLU activation function in a convolutional layer?

• To filter an image for a particular feature

• To detect important features by setting negative values to zero

• To condense the image to enhance the features

• To define the number of feature maps to create as output

Answer explanation

The ReLU (Rectified Linear Unit) activation function sets negative values to zero, thus highlighting important features by keeping only the positive values. This step is crucial for detecting features within the filtered image.

Which layer in a convolutional neural network (CNN) is responsible for the filtering step during feature extraction?

• Dense layer

• ReLU activation layer

• Convolutional layer

• Maximum pooling layer

Answer explanation

The convolutional layer is responsible for the filtering step in feature extraction. It uses kernels (filters) to scan over the image and produce weighted sums of pixel values, thus filtering the image for particular features.

What is the primary purpose of a MaxPool2D layer in a convolutional neural network?

• To detect important features by setting negative values to zero

• To filter an image for a particular feature

• To condense the feature map by retaining the most active pixels

• To increase the number of feature maps

Answer explanation

The primary purpose of a MaxPool2D layer is to condense the feature map by retaining the most active pixels within a specified patch. This helps in intensifying the features and reducing the size of the feature maps, which makes the model more efficient.

What property does maximum pooling introduce to a convolutional neural network, making it robust to variations in the position of features?

• Feature extraction

• Translation invariance

• Non-linearity

• Data augmentation

Answer explanation

Maximum pooling introduces translation invariance to a convolutional neural network. This property means that the network becomes less sensitive to small changes in the position of features within an image, allowing it to recognize the same feature regardless of its location. This makes the network more robust and reduces the amount of training data needed.

What is the primary effect of increasing the stride in a convolutional layer?

• It increases the number of feature maps created

• It results in more detailed feature extraction

• It skips over some pixels, potentially missing valuable information

• It reduces the size of the convolution kernel

Answer explanation

Increasing the stride in a convolutional layer makes the sliding window skip over some pixels in the input, which can result in missing out on important details in the feature extraction process.

What is the consequence of setting the padding parameter to 'valid' in a convolutional layer?

• The output size remains the same as the input size

• The output size shrinks compared to the input size

• The convolutional layer pads the input with zeros

• The convolutional layer uses a larger kernel

Answer explanation

When padding is set to 'valid', the convolution window stays entirely inside the input image, which leads to a reduction in the size of the output feature map compared to the input size. This can limit the depth of the network, especially with small input sizes.

What is the main purpose of using multiple convolutional blocks in a convnet?

• To reduce the number of parameters in the network

• To increase the size of the input image

• To extract increasingly complex and refined features from the image

• To decrease the training time of the network

Answer explanation

Multiple convolutional blocks allow the network to extract more complex and refined features as the data passes through each block, improving the network's ability to classify images accurately.

What is the primary purpose of data augmentation in image classification?

• To increase the model's training speed

• To artificially increase the size of the training dataset

• To reduce the number of parameters in the network

• To simplify the model architecture

Answer explanation

Data augmentation artificially increases the size of the training dataset by creating modified versions of existing images, helping the model to generalize better by seeing more varied data.