Systems and methods for partially supervised learning with momentum prototypes

A method for partially-supervised learning, the method comprising:

• obtaining, via a data interface, a training dataset comprising a set of partially-labeled data samples, each data sample having a noisy training label that belongs to a plurality of classes;
• generating, for each partially-labeled data sample, an augmented data sample by varying data content of the respective labeled partially-data sample;
• inputting the augmented data sample to the neural network model;
• generating, by a neural network model, a pseudo label corresponding to the augmented data sample based at least in part on at least one current momentum prototype associated with at least one class from the plurality of classes;
• computing a prototypical contrastive loss by comparing a normalized embedding generated from the augmented data sample with the at least one momentum prototype;
• updating the at least one momentum prototype based on the normalized embedding when the normalized embedding and the at least one momentum prototype correspond to a same class;
• and
• updating the neural network model by minimizing at least the prototypical contrastive loss.

A non-transitory processor-readable medium storing processor-executable instructions for partially-labeled learning, the instructions being processor-executable by a processor to perform operations comprising:

• obtaining a training dataset comprising a set of partially-labeled data samples, each data sample having a noisy training label that belongs to a plurality of classes;
• and
• generating, for each partially-labeled data sample, an augmented data sample by varying data content of the respective labeled partially-data sample;
• inputting the augmented data sample to a neural network model;
• generating, by the neural network model, a pseudo label corresponding to the augmented data sample based at least in part on at least one current momentum prototype associated with at least one class from the plurality of classes;
• computing a prototypical contrastive loss by comparing a normalized embedding generated from the augmented data sample with the at least one momentum prototype;
• updating the at least one momentum prototype based on the normalized embedding when the normalized embedding and the at least one momentum prototype correspond to a same class;
• and
• updating the neural network model by minimizing at least the prototypical contrastive loss.