|
I'm seeking papers on "effects of training examples' order to the classifiers' performance and accuracy". Do you know any literature on this? Jeffrey Elman has a paper related to Neural Networks(in origins of language paper),-actually this isn't a ML paper- but I couldn't find any other useful research on this. Edit: The algorithm doesn't matter much for me. But I'm curious about how the performance and the accuracy is effected. |
|
The first paper (and in my opinion a quite interesting one) that comes to mind is the one by Bengio et al. (2009) on curriculum learning. The paper by LeCun et al. 1998 explains why for stochastic gradient descend it is good to make sure that the datapoints are presented in a random order. It all depends on the kind of training algorithm you use but I assume you are thinking about stochastic gradient descend because of the reference to Elman's paper. Actually I don't have any presupposition about the algorithm but while googling, it was surprising to me not to see any papers on this topic. Because data is nearly everything in this field.
(Dec 06 '10 at 08:05)
cglr
I might be wrong but I have the impression that this is an issue that did not get the attention it deserves yet from the machine learning community.
(Dec 06 '10 at 08:09)
Philemon Brakel
|
|
By the Elman paper I'm guessing you mean Learning and development in neural networks: the importance of starting small, and that you're interested in ideas like show the simplest examples first, then more complicated, and so on? There's a 2006 paper by
Cornuejols, MACHINE LEARNING: The Necessity of Order (is order in order?) which is an overview of issues in this area, but not really a thorough survey. It's an important notion, which went out of fashion as machine learning became more and more identified with batch mode optimization. Now that online learning, lifetime learning, multi-task learning, streaming algorithms, and the like are on the ascendance I suspect we'll see more attention to this. As Philemon said, there's the more mundane issue of the importance of randomizing training data order in stochastic gradient descent. In fact I just posted a question to MA about that. Thanks for the paper Dave :) I agree with you especially for the online algorithms, data order is a crucial.
(Dec 06 '10 at 12:42)
cglr
|
