Abstract: Systems and methods for identifying and segmenting objects from images include a preprocessing module configured to adjust a size of a source image; a region-proposal module configured to propose one or more regions of interest in the size-adjusted source image; and a prediction module configured to predict a classification, bounding box coordinates, and mask. Such systems and methods may utilize end-to-end training of the modules using adversarial loss, facilitating the use of a small training set, and can be configured to process historical documents, such as large images comprising text. The preprocessing module within the systems and methods can utilize a conventional image scaler in tandem with a custom image scaler to provide a resized image suitable for GPU processing, and the region-proposal module can utilize a region-proposal network from a single-stage detection model in tandem with a two-stage detection model paradigm to capture substantially all particles in an image.

Abstract: A method and system provide for augmenting a photograph. An unlabeled photograph is obtained. A weakly augmented photograph and a strongly augmented photograph are obtained from the unlabeled photograph based on different types of data augmentation methods. The weakly augmented photograph is processed through a model to generate multiple weakly augmented photograph class predictions (with assigned probabilities). The multiple weakly augmented photograph class predictions are converted into positive pseudo-labels (indicating a presence of a class) or negative pseudo-labels (indicating absence of a class) using different fixed percentile thresholds. The strongly augmented photograph is processed through the model to generate a strongly augmented photograph class prediction. The model is trained to make the strongly augmented photograph label prediction match the positive pseudo-label via a cross-entropy loss. The trained model is then utilized to label the unlabeled photograph with multiple labels.

Abstract: Systems and methods for identifying and segmenting objects from images include a preprocessing module configured to adjust a size of a source image; a region-proposal module configured to propose one or more regions of interest in the size-adjusted source image; and a prediction module configured to predict a classification, bounding box coordinates, and mask. Such systems and methods may utilize end-to-end training of the modules using adversarial loss, facilitating the use of a small training set, and can be configured to process historical documents, such as large images comprising text. The preprocessing module within said systems and methods can utilize a conventional image scaler in tandem with a custom image scaler to provide a resized image suitable for GPU processing, and the region-proposal module can utilize a region-proposal network from a single-stage detection model in tandem with a two-stage detection model paradigm to capture substantially all particles in an image.

Abstract: Techniques for training a machine-learning (ML) model for captioning images are disclosed. A plurality of feature vectors and a plurality of visual attention maps are generated by a visual model of the ML model based on an input image. Each of the plurality of feature vectors correspond to different regions of the input image. A plurality of caption attention maps are generated by an attention model of the ML model based on the plurality of feature vectors. An attention penalty is calculated based on a comparison between the caption attention maps and the visual attention maps. A loss function is calculated based on the attention penalty. One or both of the visual model and the attention model are trained using the loss function.

Abstract: Systems and methods for training a machine learning (ML) ranking model to rank genealogy hints are described herein. One method includes retrieving a plurality of genealogy hints for a target person, where each of the plurality of genealogy hints corresponds to a genealogy item and has a hint type of a plurality of hint types. The method includes generating, for each of the plurality of genealogy hints, a feature vector having a plurality of feature values, the feature vector being included in a plurality of feature vectors. The method includes extending each of the plurality of feature vectors by at least one additional feature value based on the number of features of one or more other hint types of the plurality of hint types. The method includes training the ML ranking model using the extended plurality of feature vectors and user-provided labels.

Abstract: Hybrid machine-learning systems and methods can be used to perform automatic keyphrase extraction from input text, such as historical records. For example, a computer-implemented method for extracting keyphrases from input text can include receiving input text having a plurality of words and identifying a set of candidate phrases from the plurality of words and a score for each of the candidate phrases using one or more unsupervised machine-learning models. The method can also include identifying named entities from the set of candidate phrases using one or more supervised machine-learning models and determining an updated set of scores for at least some of the candidate phrases within the set based on the named entities identified using the supervised machine-learning model. The method can also include identifying a keyphrase from the set of candidate phrases based on the updated set of scores.

Abstract: OCR-text correction system and method embodiments are described. The OCR-text correction embodiments comprise or cooperate with a transformer-based sequence-to-sequence language model. The model is pretrained to denoise corrupted text and is fine-tuned using OCR-correction-specific examples. Text obtained at least in part through OCR is applied to the fine-tuned pretrained transformer model to detect at least one error in a subset of the text. Responsive to detecting the at least one error, the fine-tuned pretrained transformer model outputs an updated subset of the text to correct the at least one error.

Abstract: Described herein are systems, methods, and other techniques for training a generative adversarial network (GAN) to perform an image-to-image transformation for recognizing text. A pair of training images are provided to the GAN. The pair of training images include a training image containing a set of characters in handwritten form and a reference training image containing the set of characters in machine-recognizable form. The GAN includes a generator and a discriminator. The generated image is generated using the generator based on the training image. Update data is generated using the discriminator based on the generated image and the reference training image. The GAN is trained by modifying one or both of the generator and the discriminator using the update data.

Abstract: Systems and methods for training a machine learning (ML) ranking model to rank genealogy hints are described herein. One method includes retrieving a plurality of genealogy hints for a target person, where each of the plurality of genealogy hints corresponds to a genealogy item and has a hint type of a plurality of hint types. The method includes generating, for each of the plurality of genealogy hints, a feature vector having a plurality of feature values, the feature vector being included in a plurality of feature vectors. The method includes extending each of the plurality of feature vectors by at least one additional feature value based on the number of features of one or more other hint types of the plurality of hint types. The method includes training the ML ranking model using the extended plurality of feature vectors and user-provided labels.

Abstract: A computing server may continuously update a set of nodes that are addable to a data tree based on past interactions of the user with one or more nodes. The computing server may track a recently interacted set of interacted nodes with which the user has interacted within a number of past interactions. The computing server may select a pool of candidate nodes based on the recently interacted set. At least one of the candidate nodes is within a domain boundary of one of the interacted nodes that is in the recently interacted set. The domain boundary may be determined by the degree of relationship. The computing server may present one or more candidate nodes in the pool as a version of the continuously updated set of nodes. The computing server may update the pool of candidate nodes as additional interactions performed by the user updates the recently interacted set.

Abstract: Embodiments described herein relate generally to a methodology of efficient object classification within a visual medium. The methodology utilizes a first neural network to perform an attention based object localization within a visual medium to generate a visual mask. The visual mask is applied to the visual medium to generate a masked visual medium. The masked visual medium may be then fed into a second neural network to detect and classify objects within the visual medium.

Abstract: Described herein are systems, methods, and other techniques for extracting one or more keyphrases from an input text. The input text may include a plurality of words. A plurality of token-level attention matrices may be generated using a transformer-based machine learning model. The plurality of token-level attention matrices may be converted into a plurality of word-level attention matrices. A set of candidate phrases may be identified from the plurality of words based on the plurality of word-level attention matrices. The one or more keyphrases may be selected from the set of candidate phrases.

Abstract: A simplified handwriting recognition approach includes a first network comprising convolutional neural network comprising one or more convolutional layers and one or more max-pooling layers. The first network receives an input image of handwriting and outputs an embedding based thereon. A second network comprises a network of cascaded convolutional layers including one or more subnetworks configured to receive an embedding of a handwriting image and output one or more character predictions. The subnetworks are configured to downsample and flatten the embedding to a feature map and then a vector before passing the vector to a dense neural network for character prediction. Certain subnetworks are configured to concatenate an input embedding with an upsampled version of the feature map.

Abstract: Systems and methods for handwriting recognition using language modeling facilitate improved results by using a trained language model to improve results from a handwriting recognition machine learning model. The language model may be a character-based language model trained on a dataset pertinent to field values on which the handwriting recognition model is to be used. A loss prediction module may be trained with the handwriting recognition model and/or the language model and used to determine whether a prediction from the handwriting recognition model should be refined by passing the prediction through the trained language model.

Abstract: Techniques for training a machine-learning (ML) model for captioning images are disclosed. A plurality of feature vectors and a plurality of visual attention maps are generated by a visual model of the ML model based on an input image. Each of the plurality of feature vectors correspond to different regions of the input image. A plurality of caption attention maps are generated by an attention model of the ML model based on the plurality of feature vectors. An attention penalty is calculcated based on a comparison between the caption attention maps and the visual attention maps. A loss function is calculcated based on the attention penalty. One or both of the visual model and the attention model are trained using the loss function.

Abstract: Systems and methods for identifying and segmenting objects from images include a preprocessing module configured to adjust a size of a source image; a region-proposal module configured to propose one or more regions of interest in the size-adjusted source image; and a prediction module configured to predict a classification, bounding box coordinates, and mask. Such systems and methods may utilize end-to-end training of the modules using adversarial loss, facilitating the use of a small training set, and can be configured to process historical documents, such as large images comprising text. The preprocessing module within said systems and methods can utilize a conventional image scaler in tandem with a custom image scaler to provide a resized image suitable for GPU processing, and the region-proposal module can utilize a region-proposal network from a single-stage detection model in tandem with a two-stage detection model paradigm to capture substantially all particles in an image.

Abstract: Techniques for training a machine-learning (ML) model for captioning images are disclosed. A plurality of feature vectors and a plurality of visual attention maps are generated by a visual model of the ML model based on an input image. Each of the plurality of feature vectors correspond to different regions of the input image. A plurality of caption attention maps are generated by an attention model of the ML model based on the plurality of feature vectors. An attention penalty is calculated based on a comparison between the caption attention maps and the visual attention maps. A loss function is calculated based on the attention penalty. One or both of the visual model and the attention model are trained using the loss function.

Abstract: Embodiments described herein relate generally to a methodology of efficient object classification within a visual medium. The methodology utilizes a first neural network to perform an attention based object localization within a visual medium to generate a visual mask. The visual mask is applied to the visual medium to generate a masked visual medium. The masked visual medium may be then fed into a second neural network to detect and classify objects within the visual medium.

Abstract: Described herein are systems, methods, and other techniques for training a generative adversarial network (GAN) to perform an image-to-image transformation for recognizing text. A pair of training images are provided to the GAN. The pair of training images include a training image containing a set of characters in handwritten form and a reference training image containing the set of characters in machine-recognizable form. The GAN includes a generator and a discriminator. The generated image is generated using the generator based on the training image. Update data is generated using the discriminator based on the generated image and the reference training image. The GAN is trained by modifying one or both of the generator and the discriminator using the update data.

Abstract: Embodiments described herein relate generally to a methodology of efficient object classification within a visual medium. The methodology utilizes a first neural network to perform an attention based object localization within a visual medium to generate a visual mask. The visual mask is applied to the visual medium to generate a masked visual medium. The masked visual medium may be then fed into a second neural network to detect and classify objects within the visual medium.