Mạng thần kinh tích chập

A convolution layer applies a learned kernel at several spatial positions. At each position, it multiplies weights with corresponding input values and combines the products, commonly adding a bias. Deep-learning libraries often call this convolution even when the kernel is applied as an unflipped cross-correlation. Reusing a kernel across positions is weight sharing: a comparable layer does not learn a separate full kernel for every location. Local connectivity and shared weights are architectural properties, not a guarantee that a complete classifier recognizes an object equally well at every position. Padding, stride, pooling, boundary effects, and later layers all influence the result. Pooling commonly reduces spatial resolution and may reduce sensitivity to small local position changes, without guaranteeing complete translation invariance. Trained image networks often develop hierarchical features, with earlier patterns combined into more complex representations later, though every filter need not have a unique human-readable label. AlexNet is associated with the influential 2012 ImageNet classification result. Layer geometry is also practical to calculate. With a 6-by-6 input, 3-by-3 kernel, stride one, no padding, and no dilation, the kernel fits at four positions along each axis, giving a 4-by-4 output grid. Channels add another dimension. In a conventional dense convolution with three input channels and no grouping, one filter has kernel weights spanning all three channels and combines their contributions into one output feature map.