LeNet과 AlexNet: 최초의 성공

Long before todays giants, two small networks proved that learned convolutions could read images. Lets meet the pioneers.

In 1998 Yann LeCun built LeNet-5 to recognize handwritten digits on checks. It was tiny, but it worked beautifully.

LeNets pattern was clear: alternate convolution and pooling layers, then finish with a few dense layers. That recipe still echoes today.

The big idea was that LeNet learned its filters from data by backpropagation, instead of an engineer hand-coding edge detectors.

Data and compute were scarce, so progress stalled for years. The world was not yet ready to scale this idea up.

In 2012 AlexNet crushed the ImageNet contest, halving the error rate and igniting the modern deep learning boom.

AlexNet was a scaled-up LeNet: more layers, far more filters, and millions of parameters hungry for data.

AlexNet swapped slow saturating activations for the fast, simple ReLU, letting a deep net train in a reasonable time.

It ran on two GPUs in parallel. That hardware choice turned an old idea into a practical, winning system.

To stop memorizing, AlexNet added dropout in its dense layers, randomly silencing neurons during training.

Both nets are stacks of one core block. Here is a single Conv2d layer, the brick everything is built from.

LeNet showed CNNs could learn, and AlexNet proved they could win at scale. Together they launched the era you are studying. Nice start!