Convolutional Neural Networks Flashcards
What is one of the problems we face when reading image data?
When we deal with image data, the first thing we need to do to make the data readable by the network is to vectorize it to a 1-dimensional vector.
This means that we lose the local spatiality of the images - i.e. the network can’t tell where each pixel is actually located within the image.
What are Convolutional Neural Networks (CNNs)?
CNNs are a class of deep learning approaches that have been designed to preserve and extract features that consider the dependencies between the pixels.
What is the purpose of a Convolutional Layer?
A layer that is responsible for extracting features and turning them into feature maps.
What is a feature map?
A feature map is a kind of feature that reduces the size of the bitmap to improve the efficiency of the model.
How does a convolutional layer create a feature map?
A set of squares - called a ‘filter’ or ‘kernel’ - slides across the input bitmap (which may also be a convolved feature). The dot product of the area covered by that filter is used as the value for a ‘convolved feature’, also known as a feature map.
How can a convolutional layer extract ‘features’ from an image?
The contents of the filter used to convolve the image allow us to find different features. For example, if we have a 3x3 filter where the only positive values are on the bottom 3 cells, the filter will be able to detect the bottom horizontal edges of an item in the image.
What is the downside to using a convolutional layer?
Convolutional layers reduce the dimensionality of an image, thereby ‘compressing’ it, almost. Therefore, we lose some information about the image in exchange for speed and efficiency.
What is pooling?
Pooling is a post-processing compression technique applied on convolved features to further reduce their dimensionality.
What is the meaning of ‘max’ and ‘avg’ pooling?
Max pooling means that we choose the largest value within the filter. Average pooling means we choose the average of all values in the filter.
How does pooling work?
We choose a filter size - 2-dimensions - and a ‘stride’ - how far we want the filter to move in either direction. The filter slides across the image, and we store a value based on the type of pooling we use.
