Abstract: A system and methods to identify which signals are significant to an assessment of a complex machine system state in the presence of non-linearities and disjoint groupings of condition types. The system enables sub-grouping of signals corresponding to system sub-components or regions. Explanations of signal significance are derived to assist in causal analysis and operational feedback to the system is prescribed and implemented for the given condition and causality.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for training and deploying machine-learned communication. One of the methods includes: receiving an RF signal at a signal processing system for training a machine-learning network; providing the RF signal through the machine-learning network; producing an output from the machine-learning network; measuring a distance metric between the signal processing model output and a reference model output; determining modifications to the machine-learning network to reduce the distance metric between the output and the reference model output; and in response to reducing the distance metric to a value that is less than or equal to a threshold value, determining a score of the trained machine-learning network using one or more other RF signals and one or more other corresponding reference model outputs, the score indicating an a performance metric of the trained machine-learning network to perform the desired RF function.

Abstract: Implementations are described herein for analyzing existing interactive web sites to facilitate automatic engagement with those web sites, e.g., by automated assistants or via other user interfaces, with minimal effort from the hosts of those websites. For example, in various implementations, techniques described herein may be used to extract, validate, maintain, generalize, extend and/or distribute individual actions and “traces” of actions that are useable to navigate through various interactive websites. Additionally, techniques are described herein for leveraging these actions and/or traces to automate aspects of interaction with a third party website.

Abstract: Techniques are generally described for machine learning exampled-based annotation of image data. In some examples, a first machine learning model may receive a query image comprising a first depiction of an object-of-interest. In some examples, the first machine learning model may receive a target image representing a scene in which a second depiction of the object-of-interest is visually represented. In various examples, the first machine learning model may generate annotated output image data that identifies a location of the second depiction of the object-of-interest within the target image. In some examples, an object detection model may be trained based at least in part on the annotated output image data.

Abstract: Aspects of the present disclosure describe distributed fiber optic sensing (DFOS) systems, methods, and structures that employ a distributed fiber optic sensor placement procedure that advantageously provides a desirable sensor coverage over a network at minimal cost.

Abstract: Disclosed are systems and methods for predicting patient response to a treatment option. In one embodiment, the image slides from patient tissue samples are divided into patches and morphological patterns correlated with a disease outcome are labeled and given a patch-level score, based on whether the morphological patterns occur only in patients with good outcomes or patients with poor outcomes. A patient-level score can be generated based, at least partly, on the patch-level scores. Patch-level scores can identify regions of interest for targeted biomarker identification.

Abstract: The present disclosure provides a fluorescence image registration method, the method includes: acquiring a fluorescence image of a biochip; selecting a preset local region of the fluorescence image; acquiring a position of a minimum value of a sum of brightness values of pixels in a first direction and a second direction, and obtaining pixel-level registration points; dividing the pixel-level registration points into non-defective pixels and defective pixels; if the fluorescence image meets the preset standard, correcting positions of the defective pixels according to positions of the non-defective pixels; acquiring a position of a center of gravity of image points of fluorescent molecules according to a center of gravity method; fitting straight lines in the first direction and the second direction respectively according to the position of the center of gravity; and acquiring boundary points of the fluorescence image and calculating the positions of the boundary points.

Abstract: A system is provided for facilitating multi-label classification. During operation, the system maintains a set of training vectors. A respective vector represents an object and is associated with one or more labels that belong to a label set. After receiving an input vector, the system determines a similarity value between the input vector and one or more training vectors. The system further determines one or more labels associated with the input vector based on the similarity values between the input vector and the training vectors and their corresponding associated labels.

Abstract: This application describes systems and methods for generating machine learning models (MLMs). An exemplary method includes obtaining a sample and user input data characterizing a product or service. A subset of the data is selected from the sample based on sampling the sample according to the user input data. An MLM is trained by applying the data subset as training input to the MLM, thereby providing a trained MLM to emulate a customer selection process unique to the product or service. A user interface (UI) configured to receive other user input data and cause the trained MLM to execute on the other user input data, thereby testing the trained MLM, is presented. A summary of results from the execution of the trained MLM is generated and presented in the UI. The summary of results indicates a contribution to the trained MLM of each of a plurality of features.

Abstract: A mechanism is provided in a data processing system comprising at least one processor and at least one memory, the at least one memory comprising instructions executed by the at least one processor to cause the at least one processor to implement a healthcare analytics management system. A healthcare analytics development sub-system of the healthcare analytics management system develops an analytics pipeline of a set of analytics assets for a selected healthcare based on a set of business needs for a healthcare analytics client and a healthcare analytics model based on the set of analytics assets and the set of business needs. The healthcare analytics model links to the analytics pipeline. A model deployment module of a healthcare analytics operation sub-system of the healthcare analytics management system deploys the healthcare analytics model on a set of computing devices of the selected healthcare consumer.

Abstract: The embodiments of the present disclosure provide a method for predicting gas transmission loss of smart gas, implemented by a smart gas equipment management platform of an Internet of Things (IoT) system for predicting gas transmission loss of smart gas, comprising: obtaining gas flow data, gas pressure data, and ambient temperature data of a plurality of time points respectively based on gas metering devices, pressure detection devices, and temperature monitoring devices at a plurality of positions of a gas pipeline network; predicting a gas metering error based on the ambient temperature data; determining whether gas loss is abnormal loss based on the gas flow data, the gas pressure data, and the gas metering error; and in response to a determination that the gas loss is the abnormal loss, sending a warning notice.

Abstract: This disclosure aims to provide a technique for improving the accuracy of prediction. A first trained model for inferring labels for measurement data is generated based on a first data set. The first data set includes: combined data that are a combination of first measurement data, which are related to a first air conditioning apparatus, and labels set for the first measurement data; and second measurement data related to the first air conditioning apparatus.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for determining causal models for controlling environments. One of the methods includes repeatedly selecting, by a control system for the environment, control settings for the environment based on internal parameters of the control system, wherein: at least some of the control settings for the environment are selected based on a causal model, and the internal parameters include a first set of internal parameters that define a number of previously received performance metric values that are used to generate the causal model for a particular controllable element; obtaining, for each selected control setting, a performance metric value; determining that generating the causal model for the particular controllable element would result in higher system performance; and adjusting, based on the determining, the first set of internal parameters.

Abstract: Example embodiments include systems and methods for transmitting, by one or more processors coupled to memory, a request associated with a first application programming interface endpoint to an application programming interface server. The systems and methods may include retrieving, by the application programming interface server, data from one or more databases responsive to the request. The systems and methods may include transmitting, by the application programming interface server, a response to the one or more processors, the response including the data associated with at least one of independent recommendations and rankings.

Abstract: Systems and methods for automated, degradation-resistant, tuning of machine-learning (“ML”) models are provided. The systems and methods may identify an inoperative input utterance, retrieve a feature set associated with the inoperative input utterance, and generate an updated utterance-feature-intent (“UFI”) mapping based on the retrieved feature set. The systems and methods may retrain the ML model using the updated UFI mapping, and compare the accuracy of the system after the retraining and before the retraining. In a scenario where the accuracy of the system does not improve, the systems and methods may amplify the updated UFI mapping. In a scenario where the accuracy of the system does improve, the systems and methods may deploy the updated UFI mapping.

Abstract: According to one embodiment, a detection system includes an acquirer, a trainer, and a detector. The acquirer acquires first data, second data, and third data. The first data is based on an action of a first body part in a first work of a first worker having a first proficiency. The second data is based on an action of the first body part in the first work of a second worker having a second proficiency. The third data is based on an action of the first body part in the first work of a third worker. The trainer trains a recurrent neural network including a first output layer using the first data and the second data. The detector inputs the third data to the trained recurrent neural network and detects a response of the first neuron or the second neuron.

Abstract: One or more default protected attribute values may be determined for a prediction model trained based on training data including a plurality of training observations. Each of the plurality of training observations may include a respective plurality of training data values corresponding with a plurality of features. Each of the plurality of training observations may also include a respective target value. Each of the plurality of training observations may include a respective protected attribute value corresponding with a protected attribute feature. A request to determine a designated predicted target value for a designated inference observation may be received after determining the one or more default protected attribute values. The predicted target value may be determined by applying the prediction model to an inference observation and a designated default protected attribute value of the one or more default protected attribute values.

Abstract: A processor distributes, from a server, a trained supervised machine learning (ML) model and supervised and unsupervised feature information to a plurality of client devices; at each client device, trains the supervised ML model using local data to generate a local supervised ML model, constructs a local unsupervised ML model using the unsupervised feature information, and deploys the local supervised and unsupervised ML models; determining when a detection performance difference between the local supervised and unsupervised ML models reaches a threshold; identifies a proposed change to the supervised or unsupervised feature information; deploys the proposed change on one client device; responsive to determining the proposed change improves the detection performance of that client device, communicates the proposed change to a sampled set of client devices; and responsive to determining the proposed change improves the detection performance of a majority of the sampled set, communicates the proposed change to

Abstract: A method is described which includes receiving or obtaining a time series of data (S1). The method also includes storing the time series of data to a storage device without interrupting the reception of the time series of data (S2). The method also includes, for each of a plurality of base time periods, at the end of a most recently elapsed base time period (S5) and without interrupting the reception or storage of the time series of data, calculating (S6) one or more measurements based on the time series of data corresponding to the most recently elapsed base time period and updating a binary tree structure indexing the one or more measurements and the time series of data. Updating the binary tree structure includes generating a new binary tree leaf (1, 2, 4, 5) corresponding to the most recently elapsed base time period (S7).

Abstract: A thermal imaging apparatus comprising: a thermal detector device (100) comprising an array of thermal sensing pixels (102) and signal processing circuitry (104) coupled to the detector device (100). The circuitry (104) supports a background identifier (110) and a pixel classifier (112), the background identifier (110) comprising a common intensity identifier (114) and an expected background intensity calculator (116). The background identifier (110) receives pixel measurement data captured by the detector device (100) in respect of pixels of the array (102) and the common intensity identifier (114) identifies a largest number of substantially the same pixel intensity values from the pixel measurement data. The expected background intensity calculator (116) uses the largest number of substantially the same pixel intensity values to generate a model of expected background intensity levels.

Abstract: Techniques and systems are provided for training a machine learning model using different datasets to perform one or more tasks. The machine learning model can include a first sub-module configured to perform a first task and a second sub-module configured to perform a second task. The first sub-module can be selected for training using a first training dataset based on a format of the first training dataset. The first sub-module can then be trained using the first training dataset to perform the first task. The second sub-module can be selected for training using a second training dataset based on a format of the second training dataset. The second sub-module can then be trained using the second training dataset to perform the second task.

Abstract: Identifying table joins includes obtaining respective casting similarities between pairs of columns of a first table and a second table. Each pair of columns includes a first column of the first table and a second column of the second table. Ones of the pairs of columns not satisfying a casting similarity condition are discarded to obtain first join candidates. Respective string similarities for the first join candidates are obtained. Ones of the first join candidates not satisfying a string similarity condition are discarded to obtain second join candidates. Final join candidates are obtained using the respective casting similarities and the respective string similarities of the second join candidates. A selected join candidate of the final join candidates is received from a user.

Abstract: A computer-implemented method for testing failover may include: determining one or more cross-regional dependencies and traffic flow of an application in a first region of a cloud environment, wherein the one or more cross-regional dependencies include a dependency of the application in the first region of the cloud environment to one or more applications in at least one other region of the cloud environment; determining a risk score associated with performing failover of the application to a second region of the cloud environment at least based on the determined one or more cross-regional dependencies and traffic flow of the application; comparing the determined risk score with a predetermined risk score; in response to determining that the determined risk score is lower than the predetermined risk score, performing failover of the application to the second region of the cloud environment; isolating the second region of the cloud environment from the first region of the cloud environment for a predetermined

Abstract: A method may include receiving, from a client device, a reservation time and image data relating to a desired room; receiving, from a network storage device, facility data for a plurality of rooms; identifying, using an image recognition model, the desired room based on the image data and the facility data; determining an availability of the desired room based on the reservation time and the facility data; generating a first reservation option to reserve the desired room and/or a second reservation option to reserve an alternate room; transmitting, to the client device, the first reservation option and/or the second reservation option; receiving, from the client device, a user selection of the first reservation option and/or the second reservation option; and transmitting, to the network storage device, an instruction to reserve the desired room and/or the alternate room.

Abstract: An active learning framework is provided that employs a plurality of machine learning components that operate over iterations of a training phase followed by an active learning phase. In each iteration of the training phase, the machine learning components are trained from a pool of labeled observations. In the active learning phase, the machine learning components are configured to generate metrics used to control sampling of unlabeled observations for labeling such that newly labeled observations are added to a pool of labeled observations for the next iteration of the training phase. The machine learning components can include an inspection (or primary) learning component that generates a predicted label and uncertainty score for an unlabeled observation, and at least one additional component that generates a quality metric related to the unlabeled observation or the predicted label. The uncertainty score and quality metric(s) can be combined for efficient sampling of observations for labeling.

Abstract: A computer-implemented method for determining a motion trajectory for a mobile robot based on an occupancy prior indicating probabilities of presence of dynamic objects and/or individuals in a map of an environment. Occupancy priors are determined by a reward function defined by reward function parameters. The determining of the reward function parameters includes: providing semantic maps; providing training trajectories for each of semantic maps; computing a gradient as a difference between an expected mean feature count and an empirical mean feature count depending on each of the semantic maps and on each of the training trajectories, the empirical mean feature count is the average number of features accumulated over the provided training trajectories of the semantic maps, wherein the expected mean feature count is the average number of features accumulated by trajectories generated depending on the current reward function parameters; and updating the reward function parameters depending on the gradient.

Abstract: Disclosed herein are image-based object recognition method and system by and in which a learning server performs image-based object recognition based on the learning of environment variable data. The image-based object recognition method includes: receiving an image acquired through at least one camera, and segmenting the image on a per-frame basis; primarily learning labeling results for one or more objects in the image segmented on a per-frame basis; performing primary reasoning by performing object detection in the image through a model obtained as a result of the primary learning; performing data labeling based on the results of the primary reasoning, and performing secondary learning with weights allocated to respective boxes obtained by the primary reasoning and estimated as object regions; and finally detecting one or more objects in the image through a model generated as a result of the secondary learning.

Abstract: A method of mitigating quantum readout errors by stochastic matrix inversion includes performing a plurality of quantum measurements on a plurality of qubits having predetermined plurality of states to obtain a plurality of measurement outputs; selecting a model for a matrix linking the predetermined plurality of states to the plurality of measurement outputs, the model having a plurality of model parameters, wherein a number of the plurality of model parameters grows less than exponentially with a number of the plurality of qubits; training the model parameters to minimize a loss function that compares predictions of the model with the matrix; computing an inverse of the model based on the trained model parameters; and providing the computed inverse of the model to a noise prone quantum readout of the plurality of qubits to obtain a substantially noise free quantum readout.

Abstract: In some examples, a computing device may receive data from a plurality of groups of data sources. The computing device may access a plurality of data synthetization machine learning models configured for generating synthetic data. Respective ones of the data synthetization machine learning models may correspond to respective ones of the groups of data sources. The computing device generates first synthetic data by inputting, to a first data synthetization machine learning model, first data received from a first data source group, and generates second synthetic data by inputting, to a second data synthetization machine learning model, second data received from a second data source group. The computing device determines an allocation of resources based at least in part on comparing the first data and the first synthetic data with the second data and the second synthetic data.

Abstract: A graph-based clinical concept mapping algorithm maps ICD-9 (International Classification of Disease, Revision 9) and ICD-10 (International Classification of Disease, Revision 10) codes to unified Systematized Nomenclature of Medicine (SNOMED) clinical concepts to normalize longitudinal healthcare data to thereby improve tracking and the use of such data for research and commercial purposes. The graph-based clinical concept mapping algorithm advantageously combines a novel graph-based search algorithm and natural language processing to map orphan ICD codes (those without equivalents across codebases) by finding optimally relevant shared SNOMED concepts. The graph-based clinical concept mapping algorithm is further advantageously utilized to group ICD-9/10 codes into higher order, more prevalent SNOMED concepts to support clinical interpretation.

Abstract: A device may receive historic temporal data identifying events associated with a system, and may perform block bootstrapping of the hierarchical time series data, based on a hyperparameters, to generate blocks of data points of the historic time series data. The device may process the blocks of data points, with a plurality of different machine learning models, to calculate predictions, and may apply weights to the predictions to generate weighted predictions. The device may aggregate the weighted predictions to generate aggregated predictions, and may apply final weights to the aggregated predictions to generate weighted aggregated predictions. The device may aggregate the weighted aggregated predictions to generate a final prediction, and may perform one or more actions based on the final prediction.

Abstract: Techniques and apparatus for an interactive element presentation process are described. In one embodiment, for example, an apparatus may include logic operative to store a plurality of model specifications for computational models, monitor the object storage service for at least one model event using a first serverless computing service, provide the plurality of model specifications associated with the at least one model event to one of a plurality of serverless computing clusters, generate model data for each of the plurality of model specifications, store the model data for each of the plurality of computational models in the object storage service, monitor the object storage service for at least one data event associated with the model data using a second serverless computing service, cause the plurality of instances to generate a plurality of trained model specifications based on training of the plurality of computational models. Other embodiments are described.

Abstract: A dynamic power control method and system for resisting multi-user parameter biased aggregation in federated learning are provided; the method includes: (1) establishing a federated learning system model for resisting parameter biased aggregation; (2) constructing a 5 corresponding objective function based on a training purpose of the federated learning system model; (3) introducing, according to the established federated learning system model, a power control factor for resisting user biased gradient aggregation, and determining a corresponding over-the-air computation communication model; (4) processing a signal by a receiver using an incoherent energy detection method without cooperation between the receiver and a transmitter; and (5) determining a federated learning security mechanism method based on resistance against parameter biased aggregation, and completing an updating training process of the federated learning system model.

Abstract: A detection method for a malicious domain name in a domain name system (DNS) and a detection device are provided. The method includes: obtaining network connection data of an electronic device; capturing log data related to at least one domain name from the network connection data; analyzing the log data to generate at least one numerical feature related to the at least one domain name; inputting the at least one numerical feature into a multi-type prediction model, which includes a first data model and a second data model; and predicting whether a malicious domain name related to a malware or a phishing website exists in the at least one domain name by the multi-type prediction model according to the at least one numerical feature.

Abstract: A method, an electronic device, and a computer program product for processing data is disclosed. The method includes training a classification model based on a plurality of reference documents describing different objects, the trained classification model respectively associating the plurality of reference documents with the described objects. The method further includes determining from the individual words identification information that can identify the objects based on contributions of individual words in the reference documents to the association. Identification information that can identify objects in documents describing the objects may be determined, so that an identification information data set is automatically generated for training a machine learning model that is used to determine the identification information.

Abstract: A data collection system includes a classification model storing unit, a model delivery unit, a classification result storing unit, an optimum model recommendation unit, and a teacher data recording unit. The model delivery unit delivers a classification model to the user environment. The classification result storing unit classifies each of the classification models on the classification model storing unit using data with a label transmitted from the user environment as an input. The classification result storing unit stores a classification result including at least one of classification correctness or a percentage of correct answers for each input data. The optimum model recommendation unit presents an appropriate classification model for the input data based on the classification result for each of the classification models. The teacher data recording unit records the input data as teacher data or test data of the classification model.

Abstract: A machine learning model is scanned to detect actual or potential threats. The threats can be detected before execution of the machine learning model or during an isolated execution environment. The threat detection may include performing a machine learning file format check, vulnerability check, tamper check, and stenography check. The machine learning model may also be monitored in an isolated environment during an execution or runtime session. After performing a scan, the system can generate a signature based on actual, potential, or absence of detected threats.

Abstract: Churn-aware training of a classifier which reduces the difference between predictions of two different models, such as a prior generation of a classification model and a subsequent generation. A second dataset of labelled data is scored on a prior generation of a classification model, wherein the prior generation was trained on a first dataset of labelled data. A subsequent generation of a classification model is trained with the second dataset of labelled data, wherein in training of the subsequent generation, weighting of at least some of the labelled data in the second dataset, such as labelled data threat yielded an incorrect classification, is adjusted based on the score of such labelled data in the prior generation.

Abstract: In some implementations, a system may receive interaction data based on an interaction of a user with a terminal, wherein the interaction data indicates a location of the terminal. The system may determine a plurality of entities having a corresponding plurality of locations that are within a geographic area that includes the location of the terminal. The system may determine, based on the plurality of entities, one or more categories for categorizing a purpose associated with the interaction of the user with the terminal. The system may transmit, to a user device of the user, information that identifies the one or more categories. The system may receive, from the user device, information that identifies one or more user-specified categories of the one or more categories. The system may store, in the database, a record that includes the information that identifies the one or more user-specified categories.

Abstract: The present invention provides for building a knowledgebase of dependencies between Configuration Items(CIs) associated with IT computing environment. In operation, the present invention provides for mapping a plurality of Configuration Items(CI) with respective one or more actions. The present invention further provides for tracking and capturing of one or more actions performed on one or more CIs in relation to resolving an activity related to a reported CI. Further, the present invention provides for identifying dependencies between one or more CIs and the reported CI based on the captured one or more actions. Furthermore, the present invention provides for building a knowledgebase of dependencies between CIs of the computing environment based on the identified dependencies between one or more CIs and the reported CI. Yet further, the present invention provides for generating visual representations of dependencies between CIs.

Abstract: The present invention includes an embodiment that may determine an access level within an organization. The embodiment may generate a simulated scenario based on the access level. The embodiment may identify responses of the user to the generated simulated scenario. The embodiment may capture one or more input frames. The embodiment may analyze the responses and the one or more input frames and generate education for the user based on the responses and the one or more input frames.

Abstract: A comprehensive cybersecurity platform includes a cybersecurity intelligence hub, a cybersecurity sensor and one or more endpoints communicatively coupled to the cybersecurity sensor, where the platform allows for efficient scaling, analysis, and detection of malware and/or malicious activity. An endpoint includes a local data store and an agent that monitors for one or more types of events being performed on the endpoint, and performs deduplication within the local data store to identify “distinct” events. The agent provides the collected metadata of distinct events to the cybersecurity sensor which also performs deduplication within a local data store. The cybersecurity sensor sends all distinct events and/or file objects to a cybersecurity intelligence hub for analysis. The cybersecurity intelligence hub is coupled to a data management and analytics engine (DMAE) that analyzes the event and/or object using multiple services to render a verdict (e.g., benign or malicious) and issues an alert.

Abstract: A data processing method and system for automated construction, resource provisioning, data processing, feature generation, architecture selection, pipeline configuration, hyperparameter optimization, evaluation, execution, production, and deployment of machine learning models in an artificial intelligence solution development lifecycle. In accordance with various embodiments, a graphical user interface of an end user application is configured to provide a pre-configured template comprises an automated ML framework for data import, data preparation, data transformation, feature generation, algorithms selection, hyperparameters tuning, models training, evaluation, interpretation, and deployment to an end user. A configurable workflow is configured to enable a user to assemble one or more transmissible AI build/products containing one or more pipelines and/or ML models for executing one or more AI solutions.

Abstract: Viewing orientation of a display of a mobile device is changed by automatically tagging a change in a dynamic parameter of the device as intended or unintended based on a following change in a dynamic parameter of the device, the following change occurring within a predefined time period. The tagged change is used to train a detector to determine when a change in a dynamic parameter of the device implies a user intended change of viewing orientation. The trained detector is used to provide a determination of intention and a change of viewing orientation of the display is affected based on the determination.

Abstract: Techniques are provided for caching of machine learning model training parameters. One method comprises training a machine learning model using a given training dataset; and caching a parameter of the machine learning model from the training with the given training dataset. The cached parameter of the machine learning model is used for a subsequent training of the machine learning model. The caching may be performed after each of multiple iterations of the training of the machine learning model. A given cached iteration of the training of the machine learning model may be identified using a key based on: (i) a hash of the given training dataset, (ii) a hash of the machine learning model parameter, and/or (iii) hyperparameters of the machine learning model. The caching of a given iteration of the machine learning model may occur when the given cached iteration is not found in a cache memory.

Abstract: Efficient tagging of content items using content embeddings are provided. In one technique, multiple content items are stored a content embedding for content item is stored. Entity names are also stored along with an entity name embedding for each entity name. For each content item, (1) multiple content embeddings that are associated with the content item are identified; (2) a subset of the entity names is identified; and (3) for each entity name in the subset, (i) an embedding of the entity name is identified, (ii) similarity measures are generated based on the entity name embedding and the multiple content embeddings, (iii), a distribution of the similarity measures is generated, (iv) feature values are generated based on the distribution, (v) the feature values are input into a machine-learned classifier, and (vi) based on output from the classifier, it is determined whether to associate the entity name with the content item.

Abstract: One or more computing devices, systems, and/or methods for oversampling for imbalanced test data are provided. A classifier is configured to classify data points as either belonging to a first class or a second class. A determination may be made that the first class and the second class are imbalanced where a first number of data points estimated to be part of the first class is a threshold amount less than a second number of data points estimated to be part of the second class. An oversampling ratio is determined for the first class. The oversampling ratio is used to select a sample set of data points for editorial labeling, where the sampling set of data points comprises a total number of data points below a threshold amount.

Abstract: A method for improving face recognition from unseen domains by learning semantically meaningful representations is presented. The method includes obtaining face images with associated identities from a plurality of datasets, randomly selecting two datasets of the plurality of datasets to train a model, sampling batch face images and their corresponding labels, sampling triplet samples including one anchor face image, a sample face image from a same identity, and a sample face image from a different identity than that of the one anchor face image, performing a forward pass by using the samples of the selected two datasets, finding representations of the face images by using a backbone convolutional neural network (CNN), generating covariances from the representations of the face images and the backbone CNN, the covariances made in different spaces by using positive pairs and negative pairs, and employing the covariances to compute a cross-domain similarity loss function.

Abstract: A method includes receiving image data representing an environment of a robotic device from a camera on the robotic device. The method further includes applying a trained dense network to the image data to generate a set of feature values, where the trained dense network has been trained to accomplish a first robot vision task. The method additionally includes applying a trained task-specific head to the set of feature values to generate a task-specific output to accomplish a second robot vision task, where the trained task-specific head has been trained to accomplish the second robot vision task based on previously generated feature values from the trained dense network, where the second robot vision task is different from the first robot vision task. The method also includes controlling the robotic device to operate in the environment based on the task-specific output generated to accomplish the second robot vision task.

Abstract: Custom source code generation models are generated by tuning a pre-trained deep learning model by freezing the model parameters and optimizing a prefix. The tuning process is distributed across a user space and a model space where the embedding and output layers are performed in the user space and the execution of the model is performed in a model space that is isolated from the user space. The tuning process updates the embeddings of the prefix across the separate execution spaces in a manner that preserves the privacy of the data used in the tuning process.