3
$\begingroup$

I remember reading or hearing a claim that at any point in time since the publication of the MNIST dataset, it has never happened that a method not based on neural networks was the best given the state of science of that point in time.

Is this claim true?

Improve this question
$\endgroup$

1 Answer 1

Reset to default
3
$\begingroup$

To quote the relevant Wikipedia article:

The original creators of the database keep a list of some of the methods tested on it.[7] In their original paper, they use a support-vector machine to get an error rate of 0.8%

Feel free to look up that original paper, but to me the quote strongly suggests that the first record holder was a support vector machine.

Edit

As liori points out the quote is misleading: In the original paper Yann LeCun et al. actually tried a slew of methods and one version of ConvNet scored best (0.7).

But according to the MNIST-webpage, after that initial paper, DeCoste and Scholkopf, MLJ 2002, reached an error of 0.56 with a SVM. So if that webpage is complete, the claim is still false.

Improve this answer
$\endgroup$
3
  • $\begingroup$ Seems like the paper referenced in [5] (yann.lecun.com/exdb/publis/pdf/lecun-98.pdf) on Wikipedia states both the V-SVM result of 0.8 and the boosted LeNet-4 result of 0.7, so… not in this case. $\endgroup$
    – liori
    Commented Dec 16, 2016 at 23:08
  • $\begingroup$ [5] is actually this link: yann.lecun.com/exdb/mnist ,but you still seem to be right. $\endgroup$
    – BlindKungFuMaster
    Commented Dec 17, 2016 at 7:54
  • $\begingroup$ I don't understand how this answers the original question, which was supposedly asking if any other technique surpassed NNs on the MNIST task, not if before NNs there was a technique that performed better than them, as the OP wrote "since the publication of the MNIST dataset, it has never happened". $\endgroup$
    – nbro
    Commented Apr 13, 2022 at 12:32

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .