What are bottleneck layers in CNNs and why are they used in deep architectures?

Bottleneck layers reduce dimensionality before expensive convolutions, improving efficiency in deep CNNs.

Bottleneck design typically uses a 1x1 convolution to reduce channel dimensions, followed by a 3x3 convolution, and then another 1x1 convolution to restore dimensions. This reduces computational cost while preserving representational power. It is widely used in ResNet architectures to enable very deep networks without excessive computation.