konvolúciónarnet

konvolúciónarnet, commonly known as convolutional neural networks (CNNs), are a class of deep neural networks designed to process grid-like data such as images. They apply learnable filters to local regions, producing feature maps that capture spatial structure and hierarchy in the input.

Core components are convolutional layers, nonlinear activations (for example ReLU), and pooling layers. Convolutional layers apply kernels across the input with shared weights, detecting features regardless of position. This yields translation equivariance and significantly reduces the number of parameters compared to fully connected networks.

In typical CNN architectures, final layers aggregate features via fully connected nodes or global pooling for classification or other tasks. Training uses backpropagation with gradient-based optimization on large labeled datasets, often accelerated by GPUs. Regularization techniques such as batch normalization, dropout, and data augmentation help improve generalization.

History and impact: The concept traces to early work by Yann LeCun on LeNet for digit recognition in the 1990s, with major leaps from AlexNet in 2012 and subsequent networks (VGG, Inception, ResNet) expanding depth and performance on large-scale vision benchmarks. CNNs are now foundational in modern computer vision and related fields.

Variants and applications: 1D CNNs suit sequence data like audio and time series, while 3D CNNs handle video. Mobile and efficient variants use depthwise separable convolutions. Applications include image classification, object detection, semantic segmentation, medical imaging, facial recognition, and autonomous driving.