3 ms·
It would be nice if there was a little bit more motivation for using this method. You claim that the performance is superior to other methods, but there is no e
by clircle 9y ago
It would be nice if there was a little bit more motivation for using this method. You claim that the performance is superior to other methods, but there is no evidence, not even empirical.
Also what's the deal with the error plot? Is that error in the training set? Why do I care about that? I care about out of sample error.
- gbonaccorso 9y agoThank you for your comment. I'm working on some examples based on methods which are different from the ones adopted in the original paper. In that case, the comparison is done with an online perceptron and MIRA algorithm. Of course, I don't want to repeat the same experiments (that I believe correct), so I'm benchmarking in a different way and with more complex datasets. I understand this is a limitation, but I'm going to integrate with other pieces of information.
- gbonaccorso 9y agoFor the error plot, as this is an online method, I consider every sample implicitly as a test one. The goal was to show that after N correction, the algorithm converges to a weight configuration that is compatible with a specific data distribution. As I wrote, this can be a limitation (common to many online methods), because it's almost impossible to filter outliers out and the tolerance to noise is surely worse than an offline algorithm.