Sample Reweighting for Label Denoising of Neural Activity Data

Abstract

Neural decoding is a powerful technique to explore the relationship between neural activities and behaviors. It often needs massive accurately labeled data to train a model for behavior prediction. However, it is not easy to obtain accurate annotations for massive data, and the label noise is sometimes inevitable and needs to be denoised first. For annotation correction, we propose a sample reweighting method to denoise noisy labels. This method utilizes a small clean validation dataset to assign weights to the training data with label noise. A deep neural network model can be trained based on the weighted training data and the validation data. Based on the neural network model, new labels can be predicted for training data to realize the label denoising. The label denoising experiment is conducted on a functional magnetic resonance imaging dataset with class imbalance. The results show that the sample reweighting method can effectively denoise labels under different annotation qualities or noise levels for each class and it outperforms the baseline methods (validation only and semi-supervised learning). The sample reweighting method can also effectively handle the class imbalance problem. The proposed method is an effective way to tackle the noisy label problem in neural decoding.