Abstract: Embodiments described herein provide systems and methods for learning representation from unlabeled videos. Specifically, a method may comprise generating a set of strongly-augmented samples and a set of weakly-augmented samples from the unlabeled video samples; generating a set of predictive logits by inputting the set of strongly-augmented samples into a student model and a first teacher model; generating a set of artificial labels by inputting the set of weakly-augmented samples to a second teacher model that operates in parallel to the first teacher model, wherein the second teacher model shares one or more model parameters with the first teacher model; computing a loss objective based on the set of predictive logits and the set of artificial labels; updating student model parameters based on the loss objective via backpropagation; and updating the shared parameters for the first teacher model and the second teacher model based on the updated student model parameters.

Abstract: An interpolative centroid contrastive learning (ICCL) framework is disclosed for learning a more discriminative representation for tail classes. Specifically, data samples, such as natural images, are projected into a low-dimensional embedding space, and class centroids for respective classes are created as average embeddings of samples that belong to a respective class. Virtual training samples are then created by interpolating two images from two samplers: a class-agnostic sampler which returns all images from both the head class and the tail class with an equal probability, and a class-aware sampler which focuses more on tail-class images by sampling images from the tail class with a higher probability compared to images from the head class. The sampled images, e.g., images from the class-agnostic sampler and images from the class-aware sampler may be interpolated to generate interpolated images.

Abstract: The system and method are directed to a prototypical contrastive learning (PCL). The PCL explicitly encodes the hierarchical semantic structure of the dataset into the learned embedding space and prevents the network from exploiting low-level cues for solving the unsupervised learning task. The PCL includes prototypes as the latent variables to help find the maximum-likelihood estimation of the network parameters in an expectation-maximization framework. The PCL iteratively performs an E-step for finding prototypes with clustering and M-step for optimizing the network on a contrastive loss.

Abstract: Embodiments described herein provide an approach (referred to as “Co-training” mechanism throughout this disclosure) that jointly learns two representations of the training data, their class probabilities and low-dimensional embeddings. Specifically, two representations of each image sample are generated: a class probability produced by the classification head and a low-dimensional embedding produced by the projection head. The classification head is trained using memory-smoothed pseudo-labels, where pseudo-labels are smoothed by aggregating information from nearby samples in the embedding space. The projection head is trained using contrastive learning on a pseudo-label graph, where samples with similar pseudo-labels are encouraged to have similar embeddings.

Abstract: Systems and methods are provided for training object detectors of a neural network model with a mixture of label noise and bounding box noise. According to some embodiments, a learning framework is provided which jointly optimizes object labels, bounding box coordinates, and model parameters by performing alternating noise correction and model training. In some embodiments, to disentangle label noise and bounding box noise, a two-step noise correction method is employed. In some examples, the first step performs class-agnostic bounding box correction by minimizing classifier discrepancy and maximizing region objectness. In some examples, the second step uses dual detection heads for label correction and class-specific bounding box refinement.

Abstract: A learning mechanism with partially-labeled web images is provided while correcting the noise labels during the learning. Specifically, the mechanism employs a momentum prototype that represents common characteristics of a specific class. One training objective is to minimize the difference between the normalized embedding of a training image sample and the momentum prototype of the corresponding class. Meanwhile, during the training process, the momentum prototype is used to generate a pseudo label for the training image sample, which can then be used to identify and remove out of distribution (OOD) samples to correct the noisy labels from the original partially-labeled training images. The momentum prototype for each class is in turn constantly updated based on the embeddings of new training samples and their pseudo labels.

Abstract: The system and method are directed to a prototypical contrastive learning (PCL). The PCL explicitly encodes the hierarchical semantic structure of the dataset into the learned embedding space and prevents the network from exploiting low-level cues for solving the unsupervised learning task. The PCL includes prototypes as the latent variables to help find the maximum-likelihood estimation of the network parameters in an expectation-maximization framework. The PCL iteratively performs an E-step for finding prototypes with clustering and M-step for optimizing the network on a contrastive loss.

Abstract: Embodiments described herein provide systems and methods for noise-robust contrastive learning. In view of the need for a noise-robust learning system, embodiments described herein provides a contrastive learning mechanism that combats noise by learning robust representations of the noisy data samples. Specifically, the training images are projected into a low-dimensional subspace, and the geometric structure of the subspace is regularized with: (1) a consistency contrastive loss that enforces images with perturbations to have similar embeddings; and (2) a prototypical contrastive loss augmented with a predetermined learning principle, which encourages the embedding for a linearly-interpolated input to have the same linear relationship with respect to the class prototypes. The low-dimensional embeddings are also trained to reconstruct the high-dimensional features, which preserves the learned information and regularizes the classifier.

Abstract: The system and method are directed to a prototypical contrastive learning (PCL). The PCL explicitly encodes the hierarchical semantic structure of the dataset into the learned embedding space and prevents the network from exploiting low-level cues for solving the unsupervised learning task. The PCL includes prototypes as the latent variables to help find the maximum-likelihood estimation of the network parameters in an expectation-maximization framework. The PCL iteratively performs an E-step for finding prototypes with clustering and M-step for optimizing the network on a contrastive loss.

Abstract: Systems and methods are provided for training object detectors of a neural network model with a mixture of label noise and bounding box noise. According to some embodiments, a learning framework is provided which jointly optimizes object labels, bounding box coordinates, and model parameters by performing alternating noise correction and model training. In some embodiments, to disentangle label noise and bounding box noise, a two-step noise correction method is employed. In some examples, the first step performs class-agnostic bounding box correction by minimizing classifier discrepancy and maximizing region objectness. In some examples, the second step uses dual detection heads for label correction and class-specific bounding box refinement.

Abstract: A method provides learning with noisy labels. The method includes generating a first network of a machine learning model with a first set of parameter initial values, and generating a second network of the machine learning model with a second set of parameter initial values. First clean probabilities for samples in a training dataset are generated using the second network. A first labeled dataset and a first unlabeled dataset are generated from the training dataset based on the first clean probabilities. The first network is trained based on the first labeled dataset and first unlabeled dataset to update parameters of the first network.