AlexNetin

AlexNetin is a deep convolutional neural network architecture that gained prominence after winning the ImageNet Large Scale Visual Recognition Challenge (ILSVRC) in 2012. It was developed by Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton at the University of Toronto. The model demonstrated that deep learning could achieve substantial improvements on large-scale image classification tasks and helped popularize convolutional neural networks for computer vision.

AlexNetin comprises eight layers with learnable parameters: five convolutional layers followed by three fully connected layers. It processes 224x224 color images and uses ReLU activation functions. The first convolutional layer employs 11x11 kernels with a stride of 4, producing 96 feature maps, and is followed by max pooling and local response normalization. The second layer uses 5x5 kernels with 2x2 pooling and 256 maps. The next three convolutional layers use 3x3 kernels with 384, 384, and 256 maps respectively. The final three layers are fully connected, with 4096 neurons in the first two and 1000 outputs for the 1000 ImageNet classes. A softmax classifier provides the final predictions.

Training was performed on two GPUs to handle the model size and data. The network was trained on roughly 1.2 million training images with a separate validation set. Techniques to improve generalization included data augmentation through random translations and horizontal flips, as well as dropout applied to the first two fully connected layers. Stochastic gradient descent with momentum was used for optimization.

AlexNetin’s results marked a substantial advance over prior approaches for ImageNet, reducing error rates by a wide margin. Its success spurred rapid growth in deep learning for vision, influenced later architectures, and showcased the feasibility of large CNNs trained on consumer hardware.