In a convolutional neural network, what role does a kernel (filter) play?
Select an answer to reveal the explanation.
Short Explanation and Infographic
Think of a CNN kernel like a tiny flashlight sliding over an image—each position multiplies a small grid of weights with local pixels and sums them up. One kernel might light up on vertical edges; another on corners. Stack layers and you build rich feature maps. People mix this up with pooling (shrink the map) or activations (ReLU after the fact). And no, “kernel” here isn't a pretty network diagram tool. If the boss asks how the model “sees” a stop sign shape, you're talking learned filters. Exam takeaway: kernel = small convolving weight matrix for local features—not pooling, not ReLU, not a visualizer.
Full explanation below image
Full Explanation
In CNNs, a kernel (also called a filter) is a multidimensional array of learnable weights—commonly a small spatial matrix such as 3×3 or 5×5, with channels matching the input depth. During the forward pass, the kernel is convolved with the input: at each spatial location, the kernel’s weights multiply the underlying patch and the products are summed (plus bias), producing one value in an output feature map. Multiple kernels in a layer create multiple feature maps, each specialized for different local patterns. Those weights are learned by backpropagation just like dense-layer weights, but the sliding shared structure is what makes convolution efficient on images.
Early layers often learn edge- and color-sensitive detectors; deeper layers compose those responses into textures, parts, and object cues. Shared weights across positions give translation-friendly behavior and far fewer parameters than dense layers on raw pixels. Kernel size, stride, dilation, and padding control the receptive field and output resolution—key design choices in modern vision backbones from ResNet-style classifiers to detection and segmentation heads. Understanding the kernel is therefore central to reading CNN architecture diagrams and paper method sections.
The distractors confuse components of the CNN toolbox. Visualization software can display architectures but is not “the kernel” inside a convolution operation. Activation functions apply nonlinearities after linear or convolutional transforms so the network can model more than pure linear maps. Pooling layers downsample feature maps using operations like max or average and do not replace learned convolutional kernels, though both appear in classic pipelines and students often mix the vocabulary under time pressure.
In practice, libraries expose kernels as multi-dimensional tensors with shape roughly (out_channels, in_channels, height, width) for 2D convolution. Initialization, grouping, and depthwise-separable designs all still revolve around that same local filter idea. Memory aid: “convolve a little weight grid → feature map.” When a question asks what a kernel is in a CNN, answer with the small sliding weight matrix that extracts local features—not pooling, activations, or diagramming tools. If an option mentions only downsampling or only ReLU, it is describing a different layer type.