googlenet

GoogLeNet, also called Inception-v1, is the network that won the classification track of ILSVRC 2014 and proved that you could be both deeper and far more efficient than the competition at the same time. Built by a Google team (Szegedy et al.), it is a 22-layer-deep CNN assembled by stacking nine Inception modules, and its headline statistic is striking: it achieves better accuracy than AlexNet while using roughly 5 million parameters, about twelve times fewer than AlexNet's 60 million and far fewer still than VGG's 138 million. The name is a deliberate homage to LeNet, signalling a return to genuinely convolutional, efficient design rather than the parameter-heavy fully-connected tails of its contemporaries.

Its efficiency comes from two choices working together. First, the Inception modules use 1x1 convolution bottlenecks to keep the cost of multi-scale processing low. Second, GoogLeNet abandons the large fully-connected layers entirely, replacing them with global average pooling before the final classifier, which by itself removes the tens of millions of parameters that VGG and AlexNet spent on their dense tails. The result is a model that is small enough to be practical yet accurate enough to win, demonstrating that raw parameter count is not the path to accuracy and that architectural cleverness pays off.

GoogLeNet also introduced auxiliary classifiers, two extra small softmax heads attached partway up the network during training. The original motivation was to combat vanishing gradients by injecting additional gradient signal into the middle of the deep network; in practice these auxiliary heads turned out to act more as regularisers, and they were dropped at inference time. They were largely superseded once batch normalization (introduced in Inception-v2) and residual connections made very deep networks trainable by default. GoogLeNet thus stands as the first major architecture to treat computational efficiency as a first-class objective alongside accuracy, an attitude that later flowered fully in MobileNet and EfficientNet.