The technology disclosed includes a system. The system includes a spatial convolutional neural network configured to process sequencing images of clusters, and produce spatially convolved features, a filtering logic configured to select, from the spatially convolved features, a subset of spatially convolved features that contain centers of the clusters, a compression logic configured to compress the subset of spatially convolved features into a set of compressed features, a contextualization logic configured to access state information for compressed features in the set of compressed features, a temporal convolutional neural network configured to process the set of stateful compressed features, and produce temporally convolved stateful features, and a base calling logic configured to generate base calls for the clusters based on the temporally convolved stateful features.

In variants, the method for data analysis can include: determining a measurement set, optionally identifying measurements of interest, selecting measurements to composite, generating composite measurements, analyzing a batch of measurements, and optionally training a policy model.

In variants, the method for data analysis can include: determining a measurement set, optionally identifying measurements of interest, selecting measurements to composite, generating composite measurements, analyzing a batch of measurements, and optionally training a policy model.

Embodiments of this disclosure provide an action recognition device and method and an electronic device. The method includes: performing key point recognition on an object in a video frame by using a neural network to Obtain key point information and part affinity field score(s) of the object; performing key point connection according to the key point information and the part affinity field score(s); generating multiple key point connection candidates according to a result of the key point connection; for at least two of the multiple key point connection candidates, determining whether one of the at least two key point connection candidates is valid, so as to perform selection on the multiple key point connection candidates; and performing action recognition on the object according to the selected key point connection candidates. Hence, by selecting the generated key point connection candidates, accurate of action recognition in a bottom-up scheme may be improved.

Embodiments of this disclosure provide an action recognition device and method and an electronic device. The method includes: performing key point recognition on an object in a video frame by using a neural network to Obtain key point information and part affinity field score(s) of the object; performing key point connection according to the key point information and the part affinity field score(s); generating multiple key point connection candidates according to a result of the key point connection; for at least two of the multiple key point connection candidates, determining whether one of the at least two key point connection candidates is valid, so as to perform selection on the multiple key point connection candidates; and performing action recognition on the object according to the selected key point connection candidates. Hence, by selecting the generated key point connection candidates, accurate of action recognition in a bottom-up scheme may be improved.

A method for model training and deployment includes training, by a processor, a model to learn video representations with a self-supervised contrastive loss by performing progressive training in phases with an incremental number of positive instances from one or more video sequences, resetting the learning rate schedule in each of the phases, and inheriting model weights from a checkpoint from a previous training phase. The method further includes updating the trained model with the self-supervised contrastive loss given multiple positive instances obtained from Cascade K-Nearest Neighbor mining of the one or more video sequences by extracting features in different modalities to compute similarities between the one or more video sequences and selecting a top-k similar instances with features in different modalities. The method also includes fine-tuning the trained model for a downstream task. The method additionally includes deploying the trained model for a target application inference for the downstream task.

A method for model training and deployment includes training, by a processor, a model to learn video representations with a self-supervised contrastive loss by performing progressive training in phases with an incremental number of positive instances from one or more video sequences, resetting the learning rate schedule in each of the phases, and inheriting model weights from a checkpoint from a previous training phase. The method further includes updating the trained model with the self-supervised contrastive loss given multiple positive instances obtained from Cascade K-Nearest Neighbor mining of the one or more video sequences by extracting features in different modalities to compute similarities between the one or more video sequences and selecting a top-k similar instances with features in different modalities. The method also includes fine-tuning the trained model for a downstream task. The method additionally includes deploying the trained model for a target application inference for the downstream task.

The learning apparatus classifies target domain data into (N-c) classes based on unique features of the target domain data, classifies source domain data into N classes based on unique features of the source domain data, and classifies the target domain data and the source domain data into the N classes based on common features of the target domain data and the source domain data. Also, the learning apparatus calculates a first distance between the common features of the target domain data and the source domain data, and calculates a second distance between the unique features of the target domain data and the source domain data. Next, the learning apparatus updates parameters of a common feature extraction unit based on the first distance, and updates parameters of a target domain feature extraction unit and a source domain feature extraction unit based on the second distance.

The learning apparatus classifies target domain data into (N-c) classes based on unique features of the target domain data, classifies source domain data into N classes based on unique features of the source domain data, and classifies the target domain data and the source domain data into the N classes based on common features of the target domain data and the source domain data. Also, the learning apparatus calculates a first distance between the common features of the target domain data and the source domain data, and calculates a second distance between the unique features of the target domain data and the source domain data. Next, the learning apparatus updates parameters of a common feature extraction unit based on the first distance, and updates parameters of a target domain feature extraction unit and a source domain feature extraction unit based on the second distance.

Technologies are described for retrieving documents using image representations in the documents and is based on intra-image features. The identification of elements within an image representation can allow for deeper understanding of the image representation and for better relating image representations based on their intra-image features. The intra-image features present in image representations can be used in searches. Search results can further be reranked to improve search results. For example, reranking can allow search results to conform to intra-image dominant image features.