lesson_5_flashcards
What is a convolution layer?
A neural network layer designed to extract spatial features by applying filters (kernels) to localized regions of an input.
Why are convolution layers useful for images?
They exploit spatial locality and reduce parameters by sharing weights across locations, maintaining spatial structure in the output.
What is a kernel in a convolution layer?
A small matrix of weights that slides over the input to extract specific features, such as edges or textures.
What is the difference between convolution and cross-correlation?
Cross-correlation does not flip the kernel, while convolution does. In practice, deep learning frameworks implement cross-correlation for simplicity.
What is a pooling layer?
A layer that performs downsampling to reduce dimensionality and extract dominant features, commonly using max or average pooling.
What is the purpose of max pooling?
To retain the most prominent feature in a region, reducing spatial dimensions while preserving important activations.
What is the formula for the output size of a convolution layer?
((Input size - Kernel size + 2 * Padding) / Stride + 1), applied to height and width separately.
What is weight sharing in convolution layers?
A mechanism where the same kernel weights are applied across different regions of the input, reducing the number of parameters.
What is invariance in pooling layers?
The property where small translations or shifts in input features do not affect the output, aiding in robust classification.
How do convolution layers and pooling layers work together?
Convolution layers extract features, and pooling layers downsample them, reducing dimensionality and improving computational efficiency.
What are common hyperparameters in a convolution layer?
Kernel size, stride, padding, and the number of filters (output channels).
What is the relationship between depth in convolutional neural networks (CNNs) and feature abstraction?
Deeper layers capture more abstract and complex features, progressing from edges to object parts to entire objects.
What is the significance of LeNet in CNN development?
LeNet, introduced in the 1980s, was one of the first CNN architectures, designed for tasks like reading bank checks.
What is the role of automatic differentiation in CNNs?
It enables efficient gradient computation for optimizing all parameters across complex network architectures.
What is equivariance in convolution layers?
A property where translations of input features lead to corresponding translations in the output, preserving spatial relationships.
