Ahmet Iscen presents Learning with Neighbor Consistency for Noisy Labels

On 2022-04-28 11:00:00 at https://feectu.zoom.us/j/99542915897
Online event.

Recent advances in deep learning have relied on large, labelled datasets to
train high-capacity models. However, collecting large datasets in a time- and
cost-efficient manner often results in label noise. We present a method for
learning from noisy labels that leverages similarities between training
in feature space, encouraging the prediction of each example to be similar to
its nearest neighbours. Compared to training algorithms that use multiple
or distinct stages, our approach takes the form of a simple, additional
regularization term. It can be interpreted as an inductive version of the
classical, transductive label propagation algorithm. We thoroughly evaluate our
method on datasets evaluating both synthetic (CIFAR-10, CIFAR-100) and
(mini-WebVision, WebVision, Clothing1M, mini-ImageNet-Red) noise, and achieve
competitive or state-of-the-art accuracies across all of them.
Responsible person: Petr Pošík