Improved Accuracy for ECG Arrhythmia Classification using AlexNet-M-SE with Dropout and BatchNorm

Abstract - AlexNet—released by Krizhevsky, Sutskever, and Hinton in 2012—initiated the deep-learning revolution with its win in the ImageNet Large Scale Visual Recognition Challenge (ILSVRC-2012) by an unprecedented error reduction over traditional methods. The five-convolutional-layer (informally referred to as "AlexNet-5 conv") model applies max-pooling, ReLU non-linearities, local response normalization (LRN), data augmentation, dropout, and stochastic gradient descent (SGD) with momentum—every one of them trained on GPUs. This paper outlines the architecture, training pipeline, and contributions; discusses shortcomings; This paper presents future directions and open problems. Overall, AlexNet-5 is best understood as an early template whose ideas—ReLU, aggressive augmentation, dropout, and GPU-first training—remain at the core of current architectures even as the field continues to move toward deeper, normalized, and efficiency-optimized networks. Key Words: Convolutional Neural Networks, AlexNet, ReLU, Dropout, ImageNet, GPU Training, Deep Learning.