Convolutions, pooling, padding, stride, classic architectures, transfer learning, object detection and segmentation.
40 cards · basic cards · AI-written, checked twice. Edit anything.
- What is a convolution operation in a CNN?
Show answer
A mathematical operation where a small matrix (kernel/filter) slides across an input and computes element-wise products summed to produce output feature maps.- What is a kernel (or filter) in CNNs?
Show answer
A small learnable matrix that slides over the input to detect local features like edges, textures, or patterns.- What does stride control in a convolution?
Show answer
The step size by which the kernel moves across the input; larger stride reduces output spatial dimensions.- Why use padding in convolutions?
Show answer
To preserve spatial dimensions at image edges, prevent information loss at borders, and allow kernels to process edge pixels.- What is the difference between valid and same padding?
Show answer
Valid padding applies no padding and output shrinks; same padding adds zeros so output matches input spatial dimensions.- How do you calculate the output size after a convolution?
Show answer
Formula: (Input size - Kernel size + 2*Padding) / Stride + 1- What is the receptive field in CNNs?
Show answer
The region of the input that influences one output neuron; it grows exponentially as layers stack through convolution and pooling.- What are 1x1 convolutions used for?
Show answer
Dimensionality reduction of channels, feature mixing, and adding nonlinearity while reducing computation; common in bottleneck layers.- What is max pooling?
Show answer
A downsampling operation that selects the maximum value in a local window, reducing spatial dimensions while preserving dominant features.- What is average pooling?
Show answer
A downsampling operation that computes the mean of values in a local window; produces smoother output than max pooling.- Why use pooling in CNNs?
Show answer
To reduce computational complexity and parameters, achieve translation invariance, and provide implicit regularization against overfitting.- What is the key difference between max and average pooling?
Show answer
Max pooling selects the peak value in each window (preserves strongest features), while average pooling computes the mean (smoother representation).- What is LeNet-5?
Show answer
An early CNN architecture designed by Yann LeCun for digit recognition; pioneered convolutional and pooling concepts with 7 layers.- What is AlexNet?
Show answer
A deep CNN that won ImageNet 2012; pioneered ReLU activation, dropout regularization, and successful GPU training of deep networks.- What are the key characteristics of VGGNet?
Show answer
Uses stacked small 3x3 filters in depth rather than large filters; demonstrates that network depth is more important than filter size.
FlashKeepers