Excessive-dimensional Studying with Noisy Labels
Authors: Aymane El Firdoussi, Mohamed El Amine Seddik
Summary: This paper supplies theoretical insights into high-dimensional binary classification with class-conditional noisy labels. Particularly, we examine the habits of a linear classifier with a label noisiness conscious loss perform, when each the dimension of information p and the pattern dimension n are massive and comparable. Counting on random matrix idea by supposing a Gaussian combination information mannequin, the efficiency of the linear classifier when p,n→∞ is proven to converge in the direction of a restrict, involving scalar statistics of the info. Importantly, our findings present that the low-dimensional intuitions to deal with label noise don’t maintain in high-dimension, within the sense that the optimum classifier in low-dimension dramatically fails in high-dimension. Based mostly on our derivations, we design an optimized technique that’s proven to be provably extra environment friendly in dealing with noisy labels in excessive dimensions. Our theoretical conclusions are additional confirmed by experiments on actual datasets, the place we present that our optimized method outperforms the thought of baselines.
