model with ONLY 620K parameters reaching 85% validation accuracy in CIFAR10 dataset

• ~85% accuracy with only <1% of VGG-16's parameters and <6% of ResNet-18's parameters
• Great generalization. Only 0.6% difference between train/test accuracy (85.1% vs 84.5%).

• dropout in fully connected layers
• vgg style consecutive convolution in early layers
• Data augmentation (flipping, random cropping, color jittering, standard normalization)

• Resnet-style "BatchNorm layers before activation"
• small dropouts after pooling
• early exit from training based on validation loss