Consider a simple problem such as cat-classification.. what are some parameters on which an upper bound on accuracy would be expected to depend? Empirical human accuracy is of course a good metric to start with, but is there something more rigorous? The classifier could be something as simple as logistic regression or SVM, or as complex as a 100-layer deep CNN.. is there a good theoretical accuracy bound that applies in a general way? Are there tighter bounds that apply under specific circumstances?
Interesting question! The upper bound of any supervised machine/deep learning accuracy will be based on many variables but it will be constrained by the quality of the dataset. Assuming you have an almost noise-free data, theoretically your upper bound should be close to 100%! So no, i do not think there is a good theoretical accuracy bound that applies in a general way; it depends strictly on your dataset.
Interesting question! The upper bound of any supervised machine/deep learning accuracy will be based on many variables but it will be constrained by the quality of the dataset. Assuming you have an almost noise-free data, theoretically your upper bound should be close to 100%! So no, i do not think there is a good theoretical accuracy bound that applies in a general way; it depends strictly on your dataset.
As explained in any good textbook on machine learning, the theoretical limit on classification error is given by Bayes' theorem. It is theoretical in the sense that it cannot be computed accurately if the number of features is larger than two, unless a huge amount of data is available ("curse of dimensionality").
There is absolutely no reason to believe that the theoretical limit reaches 100% if the dataset is noise-free (i.e. no error on the labels). Simple counter-examples can easily be constructed. The theoretical limit depends on the amount of overlapping of the classes. Again, sorry to insist, this is textbook stuff.
- Badges
- Science topic
- Similar topics
- Data Mining
- Classification