By using smaller images, such as 64 by 64 pixels, researchers were able to train a model that achieved 93% accuracy in the ImageNet competition, surpassing the commonly used ResNet 50 model.