Famous CNN Architectures
Several landmark CNN architectures advanced the field. Understanding them gives insight into design choices that improve accuracy and efficiency.
LeNet‑5 (1998)
Created by Yann LeCun for handwritten digit recognition. Simple architecture: 2 conv + 2 pooling + 3 dense. Still works well for MNIST.
AlexNet (2012)
Won ImageNet 2012, kickstarted deep learning boom. Introduced ReLU, dropout, data augmentation, and GPU training. Architecture: 5 conv layers + 3 dense.
VGG‑16 / VGG‑19 (2014)
Very deep, uniform architecture: only 3×3 convolutions and 2×2 max pooling. Simple but parameter‑heavy (138M parameters). Known for good transfer learning features.
ResNet (Residual Networks, 2015)
Introduced skip connections (residual blocks) to allow training of very deep networks (152 layers). Overcomes vanishing gradient. Won ImageNet 2015. Still widely used.
Inception (GoogLeNet)
Uses inception modules with parallel convolutions of different kernel sizes. Efficient and deep (22 layers). Won ImageNet 2014.
MobileNet
Designed for mobile and edge devices using depthwise separable convolutions – much smaller and faster while maintaining reasonable accuracy.
Two Minute Drill
- LeNet: simple, MNIST.
- AlexNet: ReLU, dropout, GPU.
- VGG: uniform 3×3 convs, simple but large.
- ResNet: skip connections, very deep.
- MobileNet: efficient for mobile.
Need more clarification?
Drop us an email at career@quipoinfotech.com