2
$\begingroup$

I was reading this Development and validation of phenotype classifiers across multiple sites in the observational health data sciences and informatics network and came across the below paragraph. Can you please help me understand what does the highlighted term noise-tolerant learning or noisy-labeled training data` mean with a simple example and how is it useful when we don't have labels in our dataset etc? I am learning ML and your inputs would be helpful.

To address the scarcity of labeled training data, Chen et al used active learning to intelligently select training samples for labeling, demonstrating that classifier performance could be preserved with fewer samples.16 Another trend is the use of “silver standard training sets,” a semisupervised approach where training samples are labeled using an imperfect heuristic rather than by manual review.17–22 The intuition is that noise-tolerant classifiers trained on imperfectly labeled data will abstract higher order properties of the phenotype beyond the original labeling heuristic (so-called “noise-tolerant learning”23).

$\endgroup$
3
  • $\begingroup$ google gives many hits $\endgroup$ – kjetil b halvorsen Oct 14 '20 at 1:16
  • $\begingroup$ I guess this is different from what is used in the paper. $\endgroup$ – The Great Oct 14 '20 at 2:26
  • $\begingroup$ The number 23 after "noise tolerant learning" is presumably a citation. Reading the citation will probably provide context and elaboration on the sentence. $\endgroup$ – Sycorax Oct 15 '20 at 19:29
3
$\begingroup$

The paper is clear enough, noise-tolerant learning, in this case, is about label noise. In this case the labels are medical diagnosis, in part decided automatically by some algorithm. Such decisions can be noisy so the labels will be noisy.

So this is about learning algorithms which is robust against possible label noise. To give some context, the first part of the abstract of that paper is

Objective

Accurate electronic phenotyping is essential to support collaborative observational research. Supervised machine learning methods can be used to train phenotype classifiers in a high-throughput manner using imperfectly labeled data. We developed 10 phenotype classifiers using this approach and evaluated performance across multiple sites within the Observational Health Data Sciences and Informatics (OHDSI) network.

And the paper 23 from the references is General Bounds on the Number of Examples Needed for Learning Probabilistic Concepts, which indeed contains theorems about classification noise (the term used in that paper.)

$\endgroup$
2
  • $\begingroup$ Hi, Thanks for the response. Upvoted. Yes, I already read the paper. But I couldn't understand the equation component. Hence, I was trying to know whether there could be any other simple explanantion $\endgroup$ – The Great Oct 16 '20 at 5:18
  • $\begingroup$ Thanks for your help $\endgroup$ – The Great Nov 25 '20 at 0:22

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.