Abstract: Techniques are disclosed for deploying a computing resource (e.g., a service) in response to user input. A computer-implemented method can include operations of receiving (e.g., by a gateway computer of a cloud-computing environment) a request comprising an identifier for a computing component of the cloud-computing environment. The computing device receiving the request may determine whether the identifier exists in a routing table that is accessible to the computing device. If so, the request may be forwarded to the computing component. If not, the device may transmit an error code (e.g., to the user device that initiated the request) indicating the computing component is unavailable and a bootstrap request to a deployment orchestrator that is configured to deploy the requested computing component. Once deployed, the computing component may be added to a routing table such that subsequent requests can be properly routed to and processed by the computing component.

Abstract: Novel techniques are disclosed for enabling identity cloud service for virtual private label clouds (vPLCs). A vPLC is created for a reseller of a Cloud Services Provider (CSP) using CSP-provided infrastructure in a region such that the reseller can provide one or more reseller-offered cloud services to customers of the reseller. In some embodiments, the identity management may be configured with either a shared identity cloud service (IDCS) stack model or an independent IDCS stack model. In certain embodiments, two-tier vPLC-aware identity management functions are performed for resellers of the CSP and customers of the resellers.

Abstract: Data can be received that includes information corresponding to a set of users. Privacy protection protocols that apply to the data can be identified. A subset of the data can be identified as being personally identifiable information (PII) data, where the subset includes a set of PII attributes. The PII attributes can be split into categories based on a format of a data field in the PII attributes. The processed PII data can be combined with non-PII data to create processed client data. It can be determined to add noise to part of the processed PII data. An amount of noise can be determined based on the privacy protection protocols. The amount of noise can be added to part of the processed PII data to produce protected data. A machine-learning model can be trained using the protected data.

Abstract: Novel techniques of resource allocation services for virtual private label cloud (vPLC) are disclosed. A vPLC is created for a reseller of a Cloud Services Provider (CSP) using CSP-provided infrastructure in a region such that the reseller can provide one or more reseller-offered cloud services to customers of the reseller. In certain embodiments, the resource allocation services check a first-level policy and a resource database to determine whether a requested resource is allowed and available to be allocated to a vPLC associated with a reseller. The resource allocation services may further check a second-level policy and the resource database to determine whether the requested resource is allowed and available to be allocated to a customer of the reseller. In some embodiments, the resource allocation services may allocate resources for a vPLC according to a partitioning requirement.

Abstract: Novel techniques are disclosed for virtualizing a cloud infrastructure in a region provided by a cloud service provider (CSP) to allow a reseller of the CSP to provide reseller-offered cloud services using a securely isolated portion of the CSP-provided infrastructure in the region and have a direct business relationship with the reseller'customers. In certain embodiments, the CSP-provided infrastructure in a region is organized into one or more data centers. In certain embodiments, the securely isolation portion of the CSP-provided infrastructure comprises at least one compute resource or a memory resource.

Abstract: Disclosed are techniques for processing user profiles using data structures that are specialized for processing by a GPU. More particularly, the disclosed techniques relate to systems and methods for evaluating characteristics of user profiles to determine whether to offload certain user profiles to the GPU for processing or to process the user profiles locally by one or more central processing units (CPUs). Processing user profiles may include comparing the interest tags included in the user profiles with logic trees, for example, logic trees representing marketing campaigns, to identify user profiles that match the campaigns.

Abstract: Novel techniques are disclosed that enable the creation of a two-tier marketplace comprising a CSP marketplace and one or more marketplaces for virtual private label clouds (vPLCs). Each marketplace can be created and operated independently. In some embodiments, a publisher may publish a solution offering directly on a vPLC marketplace without involving the CSP marketplace. In other embodiments, a solution offering published on a marketplace may be automatically republished on another marketplace. Yet, in another embodiment, a customer subscribing to a vPLC marketplace can see a composite view of a directly published solution listing and a republished solution listing.

Abstract: Systems and methods for performing an automatic route flip are disclosed herein. The method can include receiving a request to flip a primary route and a secondary route in a communications network including at least a first host and a second host, each including a routing agent and an advertising agent. The method includes identifying the first host as having a dynamic path length and the second host as having a static path length, updating routing information in a database accessible by the first host to change the path length of the first host from a first path length to a second path length, receiving the updated routing information from the database with the routing agent of the first host, and advertising the updated routing information with the first host to at least one switch within the communications network.

Abstract: Techniques are disclosed for tuning external invocations utilizing weight-based parameter resampling. In one example, a computer system determines a plurality of samples, each sample being associated with a parameter value of a plurality of potential parameter values of a particular parameter. The computer system assigns weights to each of the parameter values, and then selects a first sample for processing via a first external invocation based on a weight of the parameter value of the first sample. The computer system then determines feedback data associated with a level of performance of the first external invocation. The computer system adjusts the weights of the parameter values of the particular parameter based on the feedback data. The computer system then selects a second sample of the plurality of samples to be processed via execution of a second external invocation based on the adjustment of weights of the parameter values.

Abstract: Novel techniques for creating service endpoints associated with different virtual private label clouds (vPLCs) for accessing a cloud service are disclosed. In certain embodiments, an endpoint management service (EMS) uses a novel architecture that enables the concurrent use of multiple vPLC-specific service endpoints with one endpoint per cloud service per vPLC to access the same cloud service running on multiple vPLC-specific resources. In some embodiments, each vPLC-specific service endpoint may be associated with a fully qualified domain name (FQDN) and an IP address.

Abstract: Techniques for predicting marketing outcomes using contrastive learning are disclosed, including: obtaining historical marketing messages; obtaining historical open rates associated respectively with the historical marketing messages; based on the historical marketing messages, generating latent space representations associated respectively with the historical marketing messages; based on the latent space representations and respective contents of the historical marketing messages, training a first machine learning model to map contents of marketing messages to corresponding latent space representations of the marketing messages; based at least on the latent space representations and the historical open rates, training a second machine learning model to map latent space representations of marketing messages to predicted open rates of the marketing messages.

Abstract: Novel techniques are disclosed for enabling customizable consoles of different virtual private label clouds (vPLCs). In some embodiments, one console server may execute multiple consoles for multiple vPLCs and CSP. In other embodiments, one console server may be dedicated to a vPLC-specific console. In certain embodiments, console customization including a customized set of console user interfaces (UIs) may be performed for each vPLC-specific console.

Abstract: The disclosed systems, methods and computer readable media relate to managing Non-Volatile Memory Express (NVMe) over Transmission Control Protocol (TCP) (NVMeOTCP) connections between a smart network interface card (smartNIC) and a block storage data plane (BSDP) of a cloud computing environment. A software agent (“agent”) executing at the smartNIC may manage a number of network paths (active and, in some cases, passive network paths). The agent may monitor the network traffic (e.g., input/output operations (IOPS)) through the paths (e.g., using established NVMeOTCP connections corresponding to the paths). If a condition is met relating to a performance threshold associated with the monitored paths, the agent may increase or decrease the number established NVMeOTCP connections to match real time network conditions.

Abstract: A secure private network connectivity system (SNCS) within a cloud service provider infrastructure (CSPI) is described that provides secure private network connectivity between external resources residing in a customer's on-premise environment and the customer's resources residing in the cloud. The SNCS provides secure private bi-directional network connectivity between external resources residing in a customer's external site representation and resources and services residing in the customer's VCN in the cloud without a user (e.g., an administrator) of the enterprise having to explicitly configure the external resources, advertise routes or set up site-to-site network connectivity.

Abstract: Aspects of the present disclosure include implementing fabric availability and synchronization (FAS) agents within a fabric network. In one example, a first FAS agent executing on a first network device may receive, from a second network device, a command to modify a configuration of a second network device. The first FAS may upgrade the configuration of the first network device based on the command from a current configuration to a new configuration. The first FAS agent increment a state identifier associated with the configuration of the first network device to a new state identifier associated with the new configuration. The first FAS agent may then transmit a control packet that includes the new state identifier. A second FAS agent executing on the second network device may receive the control packet and execute the command to update the configuration of the second network device to the new configuration.

Abstract: Novel techniques are disclosed for providing vPLC-specific metadata service including customized vPLC-specific metadata. In certain embodiments, each vPLC may generate a customized metadata using its corresponding vPLC-specific customization instructions. In some embodiments, a vPLC-specific metadata service may be performed using pre-generated customized vPLC-specific metadata, on-the-fly customized metadata, pre-generated CSP-format metadata, or combinations thereof.

Abstract: Systems and methods for route mismatch identification are disclosed herein. A method of route mismatch identification can create in cache an expected routing table based on expected routing information received by a routing agent of a host from a database accessible by each of the plurality of hosts. The method can include creating in cache an actual routing table based on actual routing information received by the routing agent of the host from an advertising agent of the host, comparing the actual routing table and the expected routing table, and taking an action based on the comparison of the actual routing table and the expected routing table.

Abstract: Techniques for facilitating connectivity to vPLCs created in a CSP-provided infrastructure in a region. Within the CSP-provided infrastructure in a region, when the destination of a packet is determined to be an endpoint associated with a particular vPLC, the packet is tagged with information related to the particular vPLC. The vPLC-related information for the particular vPLC can include, for example, a vPLC identifier identifying the particular vPLC, an identifier identifying a customer associated with the endpoint, a virtual cloud network identifier identifying a virtual cloud network (VCN) belonging to the particular vPLC and where the endpoint is part of the VCN, and other vPLC-related information. The packet is then routed or communicated within the CSP-provided infrastructure in a region along with the tagged vPLC-related information. The vPLC-related information is used as part of the connectivity and for routing of packets within the CSP-provided infrastructure in a region.

Abstract: Novel techniques for resource usage monitoring, billing, and enforcement for virtual private label clouds (vPLCs) are disclosed. In some embodiments, resource usage for a vPLC associated with a reseller is monitored at both reseller level and customer-of-reseller level using resource IDs, and stored as usage information in two levels and associated with a tenancy ID for the reseller (at the reseller level) and tenancy IDs for customers of the reseller (at the customer-of-reseller level). In some embodiments, a two-level billing process generates invoices using two-level pricing information and the generated invoices to either resellers or customers of resellers directly. In some embodiments, usage enforcement can be performed per vPLC or per customer tenancy of a reseller's customer.

Abstract: Techniques are described herein for applying access controls to logical secure elements (LSEs) running on the same secure element hardware platform. Embodiments include a firmware component that determines whether a message targeting an LSE is authorized to trigger an operation. For example, the firmware component may verify a signature of the received message using a public key, shared secret, or other access control key. Additionally or alternatively, access control policies may be defined to constrain the load of the LSEs on the SE platform hardware and/or to prioritize LSE access. For example, the access control policies may define usage thresholds, such as maximum threshold memory and/or processor utilization rates. As another example, the access controls may restrict the active time for an LSE to a threshold duration. If access constraints are violated or the message cannot be verified, then the firmware component may delay or deny the operation.

Abstract: Techniques are disclosed herein for continuous hyperparameter tuning with automatic domain weight adjustment based on periodic performance checkpoints. In one aspect, a method is provided that includes initializing a machine learning algorithm with a set of hyperparameter values and obtaining a hyperparameter objective function that is defined at least in part on a plurality of domains of a search space that is associated with the machine learning algorithm. For each trial of a hyperparameter tuning process: running the machine learning algorithm in different domains using the set of hyperparameter values, periodically checking a performance of the machine learning algorithm in the different domains based on the hyperparameter objective function; and continuing hyperparameter tuning with a new set of hyperparameter values after automatically adjusting the domain weights according to a regression status of the different domains.

Abstract: Techniques are described for transmitting metric data between tenancies. Metric data is gathered for resources within a customer tenancy of a multi-tenant environment. This metric data is sent to a service tenancy of the multi-tenant environment, where the service tenancy is separate from the customer tenancy. The metric data is validated and preprocessed within the service tenancy to make sure that all required fields (such as key-value pairs) are located within the metric data. The preprocessed metric data is then sent to a telemetry service for analysis.

Abstract: Techniques are disclosed for generating machine learning models that are insensitive to drift. A system trains a machine learning model using a divergent training dataset including synthesized data points simulating drift. The system can evaluate the machine learning models in terms of accuracy, latency, efficiency, and other metrics. Based on the evaluation, the system can select a machine learning model least susceptible to drift.

Abstract: Systems, computer-implemented methods, and computer-readable media for facilitating resource balancing based on resource capacities and resource assignments are disclosed. Electronic communications, received via interfaces, from monitoring devices to identify resource descriptions of resources may be monitored. A resource descriptions data store may be updated to associate each entity of the entities and resource capacities of each resource type of resource types. A first electronic communication, from resource-controlling systems, may be detected. Model data from a model data store may be accessed based on the identified resource descriptions. A first model may be identified based on the model data. A resources assessment corresponding may be generated based on whether a threshold is satisfied based on the first model, a first resource capacity of a first resource type, and the first electronic communication.

Abstract: Fingerprint inference of software artifacts includes receiving a request including classes, generating request fingerprints from the classes, and querying at least one index with the request fingerprints to identify a matching set of artifact versions. Fingerprint inference further includes obtaining, for each matching artifact version in the matching set of artifact versions, a count of the request fingerprints matching a indexed fingerprint related, in the at least one index, to the artifact version, and selecting a subset of the matching set of artifact versions having a count that is maximal amongst the matching set of artifact versions. Fingerprint inference further includes returning the subset of the matching set of artifact versions.

Abstract: Techniques for business-to-business (B2B) chat routing are disclosed, including: receiving, by a B2B chatbot during a chat session with a user, user input including a user-supplied business name; performing a business lookup based at least on the user-supplied business name, to obtain a canonical business name and a unique business identifier associated with the canonical business name; performing a customer relationship management (CRM) system lookup based at least on the unique business identifier, to identify a corresponding business account; routing the chat session from the B2B chatbot to a human chat agent assigned to the corresponding business account.

Abstract: In one aspect, a system receives a request to cluster a set of log records. Responsive to receiving the request, the system identifies at least one dictionary that defines a set of tokens and corresponding token weights and generates, based at least in part on the set of tokens and corresponding token weights, a set of clusters such that each cluster in the set of clusters represents a unique combination of two or more tokens from the dictionary and groups a subset of log records mapped to the unique combination of two or more tokens. The system may then perform one or more automated actions based on at least one cluster in the set of clusters.

Abstract: Methods and systems are disclosed for automatic generation of content distribution images that include receiving user input corresponding to a content-distribution operation. The user input may be parsed to identify keywords. Image data corresponding to the keywords can be identified. Image-processing operations may be executed on the image data. Executing a generative adversarial network on the processed image data, which includes: executing a first neural network on the processed-image data to generate first images that correspond to the keywords, the first images generated based on a likelihood that each image of the first images would not be detected as having been generated by the first neural network. A user interface can display the first images with second images that include images that were previously part of content-distribution operations or images that were designated by an entity as being available for content-distribution operations.

Abstract: When a request for accessing a service is received, a user object may be stored in a long-term data store, as well as in a short-term cache. The cache may be divided into a regular cache that stores full versions of the user objects, and a surrogate cache that stores compact versions of the user object. The compact version of the user object may include a field that is derived from the full user object indicating whether a subsequent request for access to a particular service should be granted. After access is granted/denied based on this value in the compact user object, the system can process an update to the full user object offline. This surrogate cache structure may be used to rapidly approve/deny requests, decoupling this procedure from the processing involved with a full user object.

Abstract: Techniques are described for enabling replication-aware resource management and task management in a cloud infrastructure for cross-region replication. In some embodiments, each replication job is associated with a set of replication-related information. In certain embodiments, the replication-aware resource management allocates resources, using a combination of various resource allocation schemes, to a fleet of replicators to allow the fleet to select replication jobs in a job queue, and perform resource scaling based on monitored performance metrics reported by the fleet. In some embodiments, the replication-aware task management enables replication job selection based on the set of replication-related information to optimize the performance of all cross-region replications in the region.

Abstract: Techniques for generating a schema transformation for application data to monitor and manage the application in a runtime environment are disclosed. A system runs an application plugin in a runtime environment to identify data generated by application modules in one or both of an application build process and an application execution process. The application plugin is a software program executed together with the application build process. The application plugin identifies a source schema associated with application data. The application plugin identifies a target schema associated with an analysis program or machine learning model. The application plugin generates a schema transformation to convert application runtime data into a target data set. The system applies the target data set to an analysis program, such as a machine learning model, to generate output analysis data associated with the application.

Abstract: Techniques discussed herein include providing a cloud computing environment in which applications are deployed by a deployment orchestrator using virtual-machine-based virtualization with a static pool of computing nodes (e.g., substrate nodes, overlay nodes) and container-based virtualization with a dynamic pool of computing nodes (e.g., nodes managed by a container orchestration platform). Components of a data plane may be used to deploy containers to micro-virtual machines. A container runtime interface (CRI) may receive a deployment request from the deployment orchestrator. A container networking interface of the data plane may configure network connections and allocate an IP address for the container. A container runtime of the data pane may generate and configure the container with the IP address and run the container within a micro-virtual machine that is compatible with the container orchestration platform.

Abstract: A computer program product, system, and computer implemented method for application-level redirect trapping and creation of NAT mapping to work with routing infrastructure for private connectivity in cloud and customer networks. The approach disclosed herein generally comprises a method of leveraging a reverse connection endpoint and IP address mapping controller to capture redirection messages from a private cloud or network (e.g., a service consumer network or a service consumer hybrid cloud). This allows at least the IP address mapping controller to manage a cloud networking infrastructure to provide for a service provider network (e.g., a public cloud) to support applications that overcome the isolation requirements of a private cloud or network to perform useful work. For example, without saddling the private cloud or network user with a heavy pre-configuration burden, the approach disclosed herein supports redirection to dynamically determined IP addresses at the private cloud or network.

Abstract: Techniques for managing composite tokens for content access requests are disclosed. A system provides a client device with a composite token to allow the client device to make subsequent requests to access content of a content provider without requiring re-authentication of the client device with each request. The composite token includes an access segment associated with permissions to access content. The composite token further includes a regeneration segment associated with permissions to invalidate the composite token and create a new composite token associated with a same user or session. The system invalidates a previous composite token and regenerates a new composite token if the access segment expires. The system requires re-authentication if the regeneration segment expires or if a composite token is received that is not the most recently-generated composite token.

Abstract: Techniques are disclosed for tuning external invocations utilizing weight-based parameter resampling. In one example, a computer system determines a plurality of samples, each sample being associated with a parameter value of a plurality of potential parameter values of a particular parameter. The computer system assigns weights to each of the parameter values, and then selects a first sample for processing via a first external invocation based on a weight of the parameter value of the first sample. The computer system then determines feedback data associated with a level of performance of the first external invocation. The computer system adjusts the weights of the parameter values of the particular parameter based on the feedback data. The computer system then selects a second sample of the plurality of samples to be processed via execution of a second external invocation based on the adjustment of weights of the parameter values.

Abstract: Techniques for UNDO and REDO operations in a computer-user interface are disclosed. The techniques enables users to configure entities for UNDO and REDO operations. The techniques also enable users to roll back individual entity to an immediate previous state in one UNDO operation and subsequently to the other previous states. Other entities are not affected by the UNDO operations to the entity.

Abstract: Server-side custom actions are associated with data fields. The data fields are associated with multiple rows of a client-side spread sheet. The server-side custom actions are defined by one or more web services. Input is received into the data fields of the client-side spread sheet. The input from the data fields is transmitted in a single request to the one or more web services. Results of the server-side custom actions are received. The results are displayed in the client-side spread sheet.

Abstract: Techniques are disclosed for automatically inferring software-defined network policies from the observed workload in a computing environment. The disclosed techniques include monitoring network traffic flow originating from network interfaces corresponding to containers that execute components of an application, recording details of a new network connection or a change in the existing network connection, obtaining information concerning the components of the application, identifying metadata for a component involved in the new network connection or the change in an existing network connection based on a comparison of the details of the new network connection or a change in the existing network connection and the information concerning the components of the application, generating a network policy for the component using at least the metadata for the component, and integrating the network policy for the component into a deployment package for the application.

Abstract: In some aspects, techniques may include monitoring a primary load of a datacenter and a reserve load of the datacenter. The primary load and reserve load can be monitored by a computing device. The primary load of the datacenter can be configured to be powered by one or more primary generator blocks having a primary capacity, and the reserve load of the datacenter can be configured to be powered by one or more reserve generator blocks having a reserve capacity. Also, the techniques may include detecting that the primary load of the datacenter exceeds the primary capacity. In addition, the techniques may include connecting the reserve generator blocks to at least one of the primary generator blocks and the primary load using a computing device switch.

Abstract: Systems and methods for a VLAN switching and routing service (VSRS) are disclosed herein. A method can include generating a table for an instance of a VSRS, which VSRS couples a first virtual layer 2 network (VLAN) with a second network. The table can contain information identifying IP addresses, MAC addresses, and virtual interface identifiers for instances within the virtual layer 2 network. The method can include receiving with the VSRS a packet from a first instance designated for delivery to a second instance within the virtual layer 2 network, identifying with the VSRS the second instance within the virtual layer 2 network for delivery of the packet based on information received with the packet and information contained within the table, and delivering the packet to the identified second instance.

Abstract: A model validation system is described that is configured to automatically validate model artifacts corresponding to models. For a model artifact being validated, the model validation system is configured to dynamically determine the validation checks to be performed for the model artifact, where the validation checks include various validation checks to be performed at the model artifact level and also for individual components included in the model artifact. The checks to be performed are dynamically determined based upon the attributes of the model artifact and of the components within the model artifact. The system is configured to generate a validation report that comprises information regarding the checks performed and the results generated from performing the various validation checks. The validation report may also include information suggesting actions for passing checks that result in a failed check, or for improving the scores of certain validation checks.

Abstract: Techniques discussed herein include providing a cloud computing environment in which applications are deployed using virtual-machine-based virtualization with a static pool of computing nodes (e.g., substrate nodes, overlay nodes) and container-based virtualization with a dynamic pool of computing nodes (e.g., nodes managed by a container orchestration platform). The control plane functionality may be invoked by a deployment orchestrator (e.g., using a client of the container orchestration platform). In some embodiments, the control plane may include a set of applications that are configured to communicate with core services for certificate generation and rotation, namespace and quota management, metric monitoring and alarming, node authentication, and cluster membership management.

Abstract: Techniques are disclosed for managing aspects of identifying and/or deploying hardware of a dedicated cloud to be hosted at a customer location (a “DRCC”). A DRCC may comprise cloud infrastructure components provided by a cloud provider but hosted by computing devices located at the customer's (a “cloud owner's”) location. Services of the central cloud-computing environment may be similarly executed at the DRCC. A number of user interfaces may be hosted within the central cloud-computing. These interfaces may be used to track deployment and region data of the DRCC. A deployment state may be transitioned from a first state to a second state based at least in part on the tracking and the deployment state may be presented at one or more user interfaces. Using the disclosed user interfaces, a user may manage the entire lifecycle of a DRCC and its corresponding hardware components.

Abstract: The present disclosure relates to systems and methods for an intelligent assistant (e.g., a chatbot) that can be used to enable a user to generate a machine learning system. Techniques can be used to automatically generate a machine learning system to assist a user. In some cases, the user may not be a software developer and may have little or no experience in either machine learning techniques or software programming. In some embodiments, a user can interact with an intelligent assistant. The interaction can be aural, textual, or through a graphical user interface. The chatbot can translate natural language inputs into a structural representation of a machine learning solution using an ontology. In this way, a user can work with artificial intelligence without being a data scientist to develop, train, refine, and compile machine learning models as stand-alone executable code.

Abstract: Techniques described herein include frameworks and models for identifying, analyzing, and addressing hangs within distributed and heterogenous computing environments. A hang detection framework may model a distributed computing environment as a complex forest of interrelated requests. The hang detection framework may generate hang graphs based upon requests that are being processed and/or waited upon within the distributed environment. For example, a node within an acyclic graph may represent an execution entity that is currently processing one or more requests. Directed edges that connect one node to another may represent requests upon which an execution entity is waiting for another execution node to fulfill. The model may be used to isolate and address the root cause of hangs within the computing environment.

Abstract: Disclosed herein is a declarative modeling paradigm for a graph model, graph-physicalization, delta load/offload, automatic generation of sub-graphs based on user entitlements, a graph-pipeline for low-code graph file formats, and machine-learning pipeline for analysis of the graph model. In a particular aspect, a computer-implemented process includes generating a graphical user interface comprising one or more tools configured to allow a user to build and visualize a graph model based on data, obtaining the data from one or more sources, receiving input from the user via the graphical user interface, where the input comprises a request to create at least two nodes representing logical entities within the data and at least one edge representing one or more relationships between the logical entities, generating the graph model based on the input from the user, analyzing the data using the graph model, and output results based on the analyzing.

Abstract: Techniques are disclosed herein for addressing catastrophic forgetting and over-generalization while training a model to transform natural language to a logical form such as a meaning representation language. The techniques include accessing training data comprising natural language examples, augmenting the training data to generate expanded training data, training a machine learning model on the expanded training data, and providing the trained machine learning model. The augmenting includes (i) generating contrastive examples by revising natural language of examples identified to have caused regression during training of a machine learning model with the training data, (ii) generating alternative examples by modifying operators of examples identified within the training data that belong to a concept that exhibits bias, or (iii) a combination of (i) and (ii).

Abstract: The present disclosure relates to systems and methods for enhancing data from disjunctive sources using a weighted interaction graph. First data about first entities can be received from a first data source. Second data about second entities at least partially different than the first entities can be received from a second data source. Relationships between each entity of the first entities and second entities can be determined, and a set of classes can be inferred from the first data and from the second data. A weighted interaction graph can be generated. The weighted interaction graph can indicate a likelihood of each entity interacting with a corresponding class. An extended set of data can be generated using the weighted interaction graph. The extended set of data can be output to facilitate communication with third entities that include the first entities and the second entities.

Abstract: Systems and methods fine-tune a pretrained machine learning model. For a model having multiple layers, an initial set of configurations is identified, each configuration establishing layers to be frozen and layers to be fine-tuned. A configuration that is optimized with respect to one or more parameters is selected, establishing a set of fine-tuning layers and a set of frozen layers. An input for the model is provided to a remote system. An output of the set of frozen layers of the model, given the provided input, is received back and locally stored. The set of fine-tuning layers of the model is loaded from the remote system. The model is fine-tuned by retrieving the locally stored output of the set of frozen layers, and updating weights associated with the set of fine-tuning layers of the machine learning model.

Abstract: Techniques are disclosed herein for training and deploying a named entity recognition model. The techniques include implementing a nested labeling scheme for named entities within the training data and then training a machine learning model on the training data The techniques further include extracting an entity hierarchy for a predicted class based on a hierarchical template associated with a composite label, where the predicted class is representative of multiple named entity classes comprising at least a parent class and a child class associated with the composite label. The techniques further include increasing the volume of training data via data mining for sequence tags in a language corpus and then training a machine learning model on the training data.