Consider that a random baseline is 10% accuracy, so 60% is pretty good, especially if the training time is functionally free to get to that accuracy. Granted t

Consider that a random baseline is 10% accuracy, so 60% is pretty good, especially if the training time is functionally free to get to that accuracy. Granted the comparison is 2500 iterations of sgd which isn't a lot. I haven't played with cifar in a while, but that's probably not a ton of savings.