Abstract: The present invention relates to an apparatus and method for generating learning data for an artificial intelligence model, which generate learning data for learning of an artificial intelligence model that detects anomalies of a plant facility, and the apparatus and method collect at least one among structured data and unstructured data, and generate a learning data set for learning of an artificial intelligence model that predicts and diagnoses anomalies of a plant facility using at least one among the structured data and the unstructured data.

Abstract: An exemplary method for detecting deepfake images and providing customized analysis comprises: receiving, from a user, a textual user inquiry regarding an image; inputting the textual inquiry and the image into a deepfake detection model, wherein the deepfake detection model comprises: an image encoder for generating a plurality of image embeddings based on the image; a text encoder for generating a plurality of textual embeddings based on the textual inquiry; one or more layers for generating a plurality of answer embeddings; and a language model for generating a textual analysis based on the plurality of answer embeddings; and outputting the textual analysis, wherein the textual analysis includes a classification result of whether the image is fake and further includes one or more visual features in the image and one or more attributes of the one or more visual features that contribute to the classification result.

Abstract: A method of imaging a sample includes acquiring one or more first images of a region of the sample at a first imaging condition with a charged particle microscope system. The one or more first images are applied to an input of a trained machine learning model to obtain a predicted image indicating atom structure probability in the region of the sample. An enhanced image indicating atom locations in the region of the sample based on the atom structure probability in the predicted image is caused to be displayed in response to obtaining the predicted image.

Abstract: Examples disclosed herein may involve a computing system that is configured to (i) identify a cannabinoid formulation for which to model efficacy for a given health condition shared by a plurality of individuals, (ii) receive respective efficacy information indicating the efficacy of the cannabinoid formulation for the plurality of individuals, (iii) receive respective genetic information for the plurality of individuals, (iv) receive respective biometric information for the plurality of individuals, (v) apply machine learning techniques to group the plurality of individuals into one or more groups based on their (a) respective efficacy information and (b) similarities in their respective genetic information and respective biometric information, and (vi) embody the one or more groups into a machine learning model that functions to (a) receive, as input data, information for a given individual and (ii) based on an evaluation of the received input data, output an efficacy prediction for the given individual.

Abstract: Methods and systems described herein are directed to automatedly generating and issuing accident data gathering recommendations for a vehicle accident occurrence. In response to a recommendation system detecting an occurrence of a vehicle accident, the system can retrieve initial accident data for the occurrence. Using the initial accident data, the system can generate accident data gathering recommendations to obtain additional accident data via mapping to one or more characteristics for the vehicle accident occurrence. The recommendations, when executed, can yield additional accident data that can be compiled with the initial accident data into a vehicle accident evidence package.

Abstract: The objective of the present invention is to provide an image classification device and a method therefor with which suitable teaching data can be created. An image classification device that carries out image classification using images which are in a class to be classified and include teaching information, and images which are in a class not to be classified and to which teaching information has not been assigned, said image classification device being characterized by being provided with: an image group input unit for receiving inputs of an image group belonging to a class to be classified and an image group belonging to a class not to be classified; and a subclassification unit for extracting a feature amount for each image in an image group, clustering the feature amounts of the images in the image group belonging to a class not to be classified, and thereby dividing the images into sub-classes.

Abstract: A system and method for intrusion detection on automotive controller area networks. The system and method can detect various CAN attacks, such as attacks that cause unintended acceleration, deactivation of vehicle's brakes, or steering the vehicle. The system and method detects changes in nuanced correlations of CAN timeseries signals and how they cluster together. The system reverse engineers CAN signals and detect masquerade attacks by analyzing timeseries extracted from raw CAN frames. Specifically, anomalies in the CAN data can be detected by computing timeseries clustering similarity using hierarchical clustering on the vehicle's CAN signals and comparing the clustering similarity across CAN captures with and without attacks.

Abstract: Synthesizing training data for training a change detection model includes receiving a patched image comprising a background image with a patch pasted into the background image. It further includes synthesizing a harmonized patched image at least in part by harmonizing the patched image using a machine learning model trained on the background image. It further includes providing as output a synthetic training sample usable to train a change detection model. The synthetic training sample includes a reference image, at least a portion of the harmonized patched image, and a corresponding mask.

Abstract: A training apparatus includes a processor which acquires an image to which identification information of a subject included in the image is attached as a first correct label the image being included in an image dataset for training and being an image for use in supervised training. The processor further attaches, to the acquired image, classification information based on a feature amount of the acquired image as a second correct label, trains a classifier using the acquired image and the second correct label attached thereto, and updates training content used in the training using the acquired image and the first correct label after training the classifier.

Abstract: Aspects of the subject technology relate to systems, methods, and computer-readable media for image classification through a two-stage classifier. Raw image data of an image gathered by a sensor associated with an AV during operation of the AV is accessed. A first stage of a two-stage classifier is applied. The first stage is trained by first raw AV data captured at varying values of one or more capture parameters associated with one or more sensors of the AV in capturing the first raw AV data. A second stage of the two-stage classifier is applied to the raw image data to generate a final classification output. The second stage of the two-stage classifier if formed by a plurality of image calibration classifiers that are trained by second raw AV data at varying values of one or more image calibration parameters.

Abstract: A perception system may be used to generate bounding boxes for objects in a vehicle scene. The perception system may receive images and feature maps corresponding to the received images. The perception system may generate multiple pluralities of windows and use the multiple pluralities of windows to enrich semantic data of the feature maps. The perception system may use the enriched semantic to generate one or more bounding boxes for objects in the vehicle scene.

Abstract: A machine learning (ML) model is trained using pairs of images. Each pair includes an image of a human face and a duplicate of the image with a computer game headset overlaid on the face using computer graphics. The ML model subsequently can be used to receive an image of a gamer wearing a headset and output a full-face image of the gamer for use in, e.g., social network settings related to the game.

Abstract: The present disclosure describes systems, robots, and methods for organizing and selecting classifiers of a library of classifiers. The classifiers of the library of classifiers can be organized in a relational model, such as a hierarchy or probability model. Instead of storing, activating, or executing the entire library of classifiers at once by a robot system, computational resource demand is reduced by executing subset of classifiers to determine context, and the determined context is used as a basis for selection of another subset of classifiers. This process can be repeated, to iteratively refine context and select more specific subsets of classifiers. A selected subset of classifiers can eventually be specific to a task to be performed by the robot system, such that the robot system can take action based on output from executing such specific classifiers.

Abstract: Implementations of this disclosure provide an anomaly detection system that automatically tunes parameters of a forecasting detector that detects anomalies in a metric time series. The anomaly detection system may implement a three-stage process where a first stage tunes a historical window parameter, a second stage tunes a current window parameter, and a third stage tunes the number of standard deviation different from historical mean required to trigger an alert. The tuned historical window length determined by the first stage may be provided to the second stage as input. Both the tuned historical window length and the tuned current window length may be provided to the third stage as input as use in determining the tuned number of standard deviations.

Abstract: A system and method for detecting tracking vehicles are provided. The method includes determining that a second vehicle is a tracking vehicle for a first vehicle by applying at least one tracking vehicle determination rule to data captured by the first vehicle with respect to the second vehicle, wherein each tracking vehicle determination rule defines a combination of parameters such that the second vehicle is determined to be the tracking vehicle when the data captured by the first vehicle includes the combination of parameters for at least one of the at least one tracking vehicle determination rule; generating a notification upon determination that the second vehicle is the tracking vehicle; and executing a resolution process based on determination that the second vehicle is the tracking vehicle.

Abstract: A visible light image generation model learning unit generates a trained visible light image generation model that generates a visible light image in a second time zone from a far-infrared image in a first time zone. The visible light image generation model learning unit includes a first learning unit that machine-learns the far-infrared image in the first time zone and a far-infrared image in the second time zone as teacher data and generates a trained first generation model that generates the far-infrared image in the second time zone from the far-infrared image in the first time zone, and a second learning unit that machine-learns the far-infrared image in the second time zone and the visible light image in the second time zone as teacher data and generates a trained second generation model that generates the visible light image in the second time zone from the far-infrared image in the second time zone.

Abstract: A mobile object control device acquires, as learning history data, driving history data obtained when a learning mobile object is operated in a risk-free environment; performs learning for imitating driving of the learning mobile object in the risk-free environment using the learning history data as training data and generates an imitation learning model; acquires, as training history data, driving history data obtained when a mobile object for generating training data is operated in the same environment as the environment in which the learning history data has been acquired; estimates whether the training history data matches the learning history data using the training history data as input to the imitation learning model and assigns a label related to risks; and infers a result for controlling a mobile object to be controlled using at least the label related to risks as training data, based on sensor information of the mobile object.

Abstract: A system to generate an image classifier and test it nearly instantaneously is described herein. Image embeddings generated by an image fingerprinting model are indexed and an associated approximate nearest neighbors (ANN) model is generated. The embeddings in the index are clustered and the clusters are labeled. Users can provide just a few images to add to the index as a labeled cluster. The ANN model is trained to receive an image embedding as input and return a score and label of the most similar identified embedding. The label may be applied if the score exceeds a threshold value. The image classifier can be tested efficiently using Leave One Out Cross Validation (“LOOCV”) to provide near-instantaneous quality indications of the image classifier to the user. Near-instantaneous indications of outliers in the provided images can also be provided to the user using a distance to the centroid calculation.

Abstract: A semi-supervised model incorporates deep feature learning and pseudo label estimation into a unified framework. The deep feature learning can include multiple convolutional neural networks (CNNs). The CNNs can be trained on available training datasets, tuned using a small amount of labeled training samples, and stored as the original models. Features are then extracted for unlabeled training samples by utilizing the original models. Multi-view clustering is used to cluster features to generate pseudo labels. Then the original models are tuned by using an updated training set that includes labeled training samples and unlabeled training samples with pseudo labels. Iterations of multi-view clustering and tuning using an updated training set can continue until the updated training set is stable.

Abstract: A method may include executing a neural network to extract a first plurality of features from a plurality of first training images and a second plurality of features from a second training image; generating a model comprising a first image performance score for each of the plurality of first training images and a feature weight for each feature, the feature weight for each feature of the first plurality of features calculated based on an impact of a variation in the feature on first image performance scores of the plurality of first training images; training the model by adjusting the impact of a variation of each of a first set of features that correspond to the second plurality of features; executing the model using a third set of features from a candidate image to generate a candidate image performance score; and generating a record identifying the candidate image performance score.

Abstract: A method for lifelong machine learning using boosting includes receiving a new task and a learning sample for the new task. A distribution of weights is learned over the learning sample using previously learned classifiers from old tasks. A set of task-specific classifiers are learned for the new task using a boosting algorithm and the distribution of weights over the learning sample, whereby the distribution of weights over the learning sample is updated using the task-specific classifiers for the new task.

Abstract: A multiple scenario-oriented item retrieval method and system. The method includes the steps of extracting, by Hashing learning, image features from an image training set to train a pre-built item retrieval model; when an image is in a scenario of hard samples, introducing an adaptive similarity matrix, optimizing the similarity matrix by an image transfer matrix, constructing an adaptive similarity matrix objective function in combination with an image category label; constructing a loss quantization objective function between the image and a Hash code according to the image transfer matrix; when the image is in a scenario of zero samples, introducing an asymmetric similarity matrix, constructing an objective function by taking the image category label as supervisory information in combination with equilibrium and decorrelation constraints of the Hash code; and training the item retrieval model based on the above objective function to obtain a retrieved result of a target item image.

Abstract: Systems and methods are provided for multi-modal test-time adaptation. The method includes inputting a digital image into a pre-trained Camera Intra-modal Pseudo-label Generator, and inputting a point cloud set into a pre-trained Lidar Intra-modal Pseudo-label Generator. The method further includes applying a fast 2-dimension (2D) model, and a slow 2D model, to the inputted digital image to apply pseudo-labels, and applying a fast 3-dimension (3D) model, and a slow 3D model, to the inputted point cloud set to apply pseudo-labels. The method further includes fusing pseudo-label predictions from the fast models and the slow models through an Inter-modal Pseudo-label Refinement module to obtain robust pseudo labels, and measuring a prediction consistency for the pseudo-labels.

Abstract: Embodiments of the present disclosure relate to a method, device and computer readable storage medium of scene text detection. In the method, a first visual representation of a first image is generated with an image encoding process. A first textual representation of a first text unit in the first image is generated with a text encoding process based on a first plurality of symbols obtained by masking a first symbol of a plurality of symbols in the first text unit. A first prediction of the masked first symbol is determined with a decoding process based on the first visual and textual representations. At least the image encoding process is updating according to at least a first training objective to increase at least similarity of the first prediction and the masked first symbol.

Abstract: The second multi-dimensional feature vectors 92a of sample image data 34a having instruction signals that are converted by a feature converter 27 are read in (Step S10), two-dimensional graph data for model 36a is generated based on the read second multi-dimensional feature vectors 92a to be stored (Step S12), two-dimensional model graphs Og and Ng are generated based on the generated two-dimensional graph data for model 36a, to be displayed on the window 62 (Step S14). The second multi-dimensional feature vectors 92a are indicators appropriate for visualization of the trained state (individuality) of a trained model 35. Thus, it is possible to visually check and evaluate whether the trained model 35 is in an appropriately trained state (individuality) or not.

Abstract: Disclosed herein are systems and methods for using a robotic surgical system comprising a GUI and a robotic arm.

Abstract: The present disclosure describes systems, robots, and methods for organizing and selecting classifiers of a library of classifiers. The classifiers of the library of classifiers can be organized in a relational model, such as a hierarchy or probability model. Instead of storing, activating, or executing the entire library of classifiers at once by a robot system, computational resource demand is reduced by executing subset of classifiers to determine context, and the determined context is used as a basis for selection of another subset of classifiers. This process can be repeated, to iteratively refine context and select more specific subsets of classifiers. A selected subset of classifiers can eventually be specific to a task to be performed by the robot system, such that the robot system can take action based on output from executing such specific classifiers.

Abstract: A computer-implemented method can include a training phase and a hemorrhage detection phase. The training phase can include: receiving a first plurality of frames from at least one original computed tomography (CT) scan of a target subject, wherein each frame may or may not include a visual indication of a hemorrhage, and further wherein each frame including a visual indication of a hemorrhage has at least one label associated therewith; and using a fully convolutional neural network (FCN) to train a model by determining, for each of the first plurality of frames, whether at least one sub-portion of the frame includes a visual indication of a hemorrhage and classifying the sub-portion of the frame based on the determining.

Abstract: Provided are computing systems, methods, and platforms for generating training and testing data for machine-learning models. The operations can include receiving signal extraction information that has instructions to query a data store. Additionally, the operations can include accessing, using Structured Query Language (SQL) code generated based on the signal extraction information, raw data from the data store. Moreover, the operations can include processing the raw data using signal configuration information to generate a plurality of signals. The signal configuration information can have instructions on how to generate the plurality of signals from the raw data. Furthermore, the operations can include joining, using SQL code, the plurality of signals with a first label source to generate training data and testing data. Subsequently, the operations can include processing the training data and the testing data to generate the input data. The input data being an ingestible for a machine-learning pipeline.

Abstract: The present disclosure describes approaches for evaluating interest points for localization uses based on a repeatability of the detection of the interest point in images capturing a scene that includes the interest point. The repeatability of interest points is determined by using a trained repeatability model. The repeatability model is trained by analyzing a time series of images of a scene and determining repeatability functions for each interest point in the scene. The repeatability function is determined by identifying which images in the time series of images allowed for the detection of the interest point by an interest point detection model.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for object detection. In one aspect, a method comprises: obtaining: (i) an image, and (ii) a set of one or more query embeddings, wherein each query embedding represents a respective category of object; processing the image and the set of query embeddings using an object detection neural network to generate object detection data for the image, comprising: processing the image using an image encoding subnetwork of the object detection neural network to generate a set of object embeddings; processing each object embedding using a localization subnetwork to generate localization data defining a corresponding region of the image; and processing: (i) the set of object embeddings, and (ii) the set of query embeddings, using a classification subnetwork to generate, for each object embedding, a respective classification score distribution over the set of query embeddings.

Abstract: A recording medium determination apparatus includes an image data acquisition unit configured to acquire image data obtained by capturing an image of a predetermined area in a recording medium, a first extraction unit configured to extract a first feature amount by processing the image data using a first parameter, a second extraction unit configured to extract a second feature amount by processing the image data using a second parameter different from the first parameter, and a determination unit configured to determine a type of the recording medium based on the first feature amount and the second feature amount.

Abstract: At least one processor generates a crowd state image as an image in which a person image corresponding to a person state is synthesized with previously-prepared image at a predetermined size. The previously-prepared image is an background image that include no person. The at least one processor specifies a training label for the crowd state image. The at least one processor outputs a pair of crowd state image and training label.

Abstract: Systems, methods and non-transitory computer readable media for generating content using a generative model without relying on selected training examples are provided. An input indicative of a desire to generate a new content using a generative model may be received. The generative model may be a result of training a machine learning model using a plurality of training examples. Each training example of the plurality of training examples may be associated with a respective content. Further, an indication of a particular subgroup of at least one but not all of the plurality of training examples may be obtained. Based on the indication, the input and the generative model may be used to generate the new content, abstaining from basing the generation of the new content on any training example included in the particular subgroup. The new content may be provided.

Abstract: An information processing apparatus outputs operational information as information regarding a first operation and a second operation performed by an operator on an object as an operation target, the information processing apparatus including a scene estimator, a chunk estimator, a transition destination suggestion unit, and an output unit. The scene estimator obtains first images as images of a scene in a state where the operator performs the first operation and the second operation, and estimates the scene by using a first learned model describing an association between the first image and a scene ID that uniquely indicates the scene. The chunk estimator obtains second images as images of an object of the first operation and the second operation, and estimates a chunk by using one of a plurality of second learned models that store an association between the second image and one or a plurality of meta IDs for chunk.

Abstract: An image processing method includes steps of acquiring first model output generated based on a captured image by a first machine learning model, acquiring second model output generated based on the captured image by a second machine learning model which is different from the first machine learning model, and generating an estimated image by using the first model output and the second model output, based on a comparison based on the second model output and one of the captured image and first model output.

Abstract: Methods, systems, and apparatus for receiving a request for a damage propensity score for a parcel, receiving imaging data for the parcel, wherein the imaging data comprises street-view imaging data of the parcel, extracting, by a machine-learned model including multiple classifiers, characteristics of vulnerability features for the parcel from the imaging data, determining, by the machine-learned model and from the characteristics of the vulnerability features, a damage propensity score for the parcel, and providing a representation of the damage propensity score for display.

Abstract: A pipeline-based machine learning method includes: determining a plurality of target components from candidate components configured to construct a machine learning model; performing standardization processing on input data and output data of the plurality of target components to obtain a plurality of standardized components respectively corresponding to the plurality of target components; assembling, based on a connection relationship between the plurality of standardized components, the plurality of standardized components into a pipeline corresponding to the machine learning model; performing scheduling processing on the plurality of standardized components included in the pipeline, to obtain a scheduling result; and executing, based on THE scheduling result, a machine learning task corresponding to the machine learning model.

Abstract: A method and system may survey a property using aerial images captured from an unmanned aerial vehicle (UAV), a manned aerial vehicle (MAV) or from a satellite device. The method may include identifying a commercial property for a UAV to perform surveillance, and directing the UAV to hover over the commercial property and capture aerial images at predetermined time intervals. Furthermore, the method may include receiving the aerial images of the commercial property captured at the predetermined time intervals, detecting a surveillance event at the commercial property, generating a surveillance alert, and transmitting the surveillance alert to an electronic device associated with an owner of the commercial property.

Abstract: Embodiments described herein provide for training a machine learning model for automatic organ segmentation. A processor executes a machine learning model using an image to output at least one predicted organ label for a plurality of pixels of the image. Upon transmitting the at least one predicted organ label to a correction computing device, the processor receives one or more image fragments identifying corrections to the at least one predicted organ label. Upon transmitting the one or more image fragments and the image to a plurality of reviewer computing devices, the processor receives a plurality of inputs indicating whether the one or more image fragments are correct. When a number of inputs indicating an image fragment of the image fragments is correct exceeds a threshold, the processor aggregates the image fragment into a training data set. The processor trains the machine learning model with the training data set.

Abstract: An auto-tagging engine receives a training set of data comprising documents including a set of tagged fields with each tagged field corresponding to a portion of the document. The auto-tagging engine trains a machine learned model using the training set of data. The trained machine learned model, when applied to a target document in a document management environment, identifies portions of the target document each corresponding to fields of the target document. For each field of the target document, the auto-tagging engine identifies text of the target document associated with the identified portions of the target document corresponding to fields. Natural language processing is performed on the identified text in order to identify field types for the fields. The target document is automatically modified to include a tag identifying the portion of the target document corresponding to each field and identifying a field type of the field.

Abstract: Provided is an image analysis unit which analyzes a captured image of a camera mounted on a mobile device, executes object identification of an image, and sets a label as an identification result to an image region; a low-confidence region extraction unit which extracts a region with low confidence of object identification from an image analysis result; and a label updating unit which updates a label of the low-confidence region on the basis of information received via a communication unit. The label updating unit updates a label in a case where a matching rate between an object region analyzed from information received via the communication unit and the low-confidence region is equal to or greater than a specified threshold.

Abstract: A method and system for detecting a typical object of a transmission line based on UAV federated learning. The method includes: determining a detection model for a typical object of a transmission line by YOLOv3 object detection algorithm according to a prior database for the typical object; dividing a UAV network into multiple federated learning units; acquiring pictures, taken by the UAV network, of the typical object and tags corresponding to each picture to determine a training database; training, based on Horovod framework and FATE federated learning framework, each federated learning unit according to the training database and the detection model for the typical object, and determining the trained UAV network according to the trained federated learning unit; and determining, by the trained UAV network, the typical object in each picture. A congestion of communication links is avoided, thereby improving detection efficiency.

Abstract: An abnormality detection apparatus (100) includes an acquisition unit (110) that acquires input data, an abnormality degree computation unit (120) that has a discriminative model for computing an abnormality degree of input data, inputs the acquired input data to the discriminative model, and thereby computes an abnormality degree of the input data, a normality degree computation unit (130) that has a normality model for computing a normality degree of input data, inputs the input data to the normality model, and thereby computes a normality degree of the input data, a determination unit (140) that has a determination model for performing determination relating to an abnormality level of input data, inputs the abnormality degree and the normality degree to the determination model, and thereby performs determination relating to an abnormality level of the input data, and an output unit (150) that outputs output information based on a result of the determination.

Abstract: Methods and systems for providing an interactive image scene graph pattern search are provided. A user is provide with an image having a plurality of selectable segmented regions therein. The user selects one or more of the segmented regions to build a query graph. Via a graph neural network, matching target graphs are retrieved that contain the query graph from a target graph database. Each matching target graph has matching target nodes that match with the query nodes of the query graph. Matching target images from an image database are associated with the matching target graphs. Embeddings of each of the query nodes and the matching target nodes are extracted. A comparison of the embeddings of each query node with the embeddings of each matching target node is performed. The user interface displays the matching target images that are associated with the matching target graphs.

Abstract: Techniques for synthetic data generation in computer-based reasoning systems are discussed and include receiving a request for generation of synthetic data based on a set of training data cases. One or more focal training data cases are determined. For undetermined features (either all of them or those that are not subject to conditions), a value for the feature is determined based on the focal cases. In some embodiments, the generated synthetic data may be checked for similarity against the training data, and if similarity conditions are met, it may be modified (e.g., resampled), removed, and/or replaced.

Abstract: Systems and methods for analyzing image data to identify cabinet products are disclosed. A computer-implemented method may include receiving, from an electronic device via a network connection, at least one digital image depicting a cabinet. The method also may include analyzing, by one or more processors, the at least one digital image to determine a first set of characteristics of the cabinet. Additionally, the method may include accessing, by the one or more processors from memory, a second set of characteristics corresponding to a plurality of cabinet products and comparing the first set of characteristics to the second set of characteristics to identify a cabinet product of the plurality of cabinet products that matches the cabinet. Further, the method may include transmitting, to the electronic device via the network connection, an indication of the cabinet product.

Abstract: A method for context based lane prediction, the method may include obtaining sensed information regarding an environment of the vehicle; providing the sensed information to a second trained machine learning process; and locating one or more lane boundaries by the second trained machine learning process. The second trained machine learning process is generated by: performing a self-supervised training process, using a first dataset, of a first machine learning process to provide a first trained machine learning process; wherein the first trained machine learning process comprises a first encoder portion and a first decoder portion; replacing the first decoder portion by a second decoder portion to provide a second machine learning process; and performing an additional training process, using a second dataset that is associated with lane boundary metadata, of the second machine learning process to provide a second trained machine learning process.

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: A system including one or more processors and one or more non-transitory computer-readable media storing computing instructions configured to run on the one or more processors and perform generating one or more item relational graphs for one or more items based on historical user purchases; transforming, using spectral filtering, the one or more item relational graphs into one or more frequency signals to remove noise from the one or more frequency signals; constructing, using a machine learning model, one or more item pair label classifications for one or more item pairs of the one or more items; generating a respective similarity score for each of the one or more item pairs; outputting a top k results for the one or more item pairs ranked by the respective similarity scores; and re-ranking, using a re-ranking algorithm, the top k results of the one or more item pairs based on a user preference for display on a user interface of an electronic device of a user. Other embodiments are disclosed.