Convolutional networks dominate state-of-the-art computer vision tasks. Deep CNNs mainstream since 2014 yield benchmark gains. Larger models boost quality with sufficient data but demand efficiency for mobile and big-data use. Factorized convolutions scale computation efficiently. Aggressive regularization curbs overfitting. Single model hits 21.2% top-1 and 5.6% top-5 error on ILSVRC 2012 validation. Achieves this with 5 billion multiply-adds per inference. Uses fewer than 25 million parameters. Ensemble of 4 models with multi-crop evaluation scores 3.5% top-5 validation error. Test set top-5 error is 3.6%. Validation top-1 error is 17.3%. Methods surpass prior state-of-the-art.