Abstract: A data processing system implements analyzing input data from analytics platform using a large language model to cause the model to analyze the metadata and generate a list of causal confounders that may influence a causal relationship between an independent variable and an output being studied; generating a plurality of causal hypothesis scenarios testing how one or more causal confounders of the list of causal confounders influence a relationship between the independent variable and the output being studied; executing the plurality of causal hypothesis scenarios to obtain causal inference results; performing statistical testing on the causal results to determine whether each confounder of the list of causal confounders satisfies a significance threshold; pruning the list of causal confounders to remove confounders that do not satisfy the significance threshold; and generating a report that identifies each of the confounders and includes a description of a significance of each of the confounders.

Abstract: A facility for managing information about document-related activities is described. In some cases, the facility uses particular kinds of structures to represent, in a document graph, document-related activities performed by particular users. In some cases, the facility uses these structures to enable one user to subscribe to the document-related activities performed by another user. In some cases, the facility uses these structures to aggregate document-related activities performed by users in a group of users, such as by aggregating topics that are addressed by documents that are the subject of these document-related activities. In some cases, the facility uses particular kinds of structures to represent, in a document graph, tasks that certain users either can help with or need help with. In some cases, the facility uses these structures to match users that can help with a task with users that need help with a task.

Abstract: An image restoration machine-trained model (“model”) removes degradation from an input image that is associated with one or more degradation types. The model has an encoder-decoder framework that includes encoders and decoders. Each encoder and decoder processes feature information in a channel attention operation and a spatial attention operation. The model performs the channel attention operation using at least two different channel attention techniques that consume different amounts of resources. The model performs the spatial attention operation by applying different spatial attention techniques to different types of content in the input image. The feature information passed to the decoders includes degradation information that describes the types degradation in the input image. A training system updates the parameters of the model upon the introduction of a new degradation type by fine-tuning an existing version of the model, rather than training the model from scratch.

Abstract: An optic comprises an optical waveguide, a series of anti-reflective laminae, and an exit grating. The waveguide supports total internal reflection of display light from the surface. The series of anti-reflective laminae are formed on the optical waveguide and arranged parallel to the surface. The exit grating is formed on the series of anti-reflective laminae. The exit grating includes a two-dimensional matrix of replicated, nanometer-scale features distributed parallel to the surface.

Abstract: An optic comprises an optical waveguide and an exit grating configured to out-couple light from the optical waveguide. The optical waveguide has a surface and supports total internal reflection of display light from the surface. The exit grating includes a two-dimensional matrix of replicated, nanometer-scale features distributed parallel to the surface.

Abstract: The description relates to adaptive data storage management. An example manages data within a database system using dynamically adjustable mini-pages within a buffer pool. The example involves receiving a request for a data item, adjusting mini-page sizes based on the request, and determining the data item's presence within these mini-pages. Upon locating the item, the request is executed on the relevant mini-page. The example also updates a mapping table to correlate mini-page identifiers with their locations in the database, and selectively caches or evicts the data item based on probabilities and access patterns.

Abstract: In one example of the present technology, a method includes sending, by a CPU core of a processor, a direct memory access (DMA) stream setup command to an input/output (I/O) device, the DMA stream setup command comprising a virtual address; and sending, from the CPU core of the processor, a prefetch message comprising the virtual address to an input/output memory management unit (IOMMU) to cause the IOMMU to: prefetch a translation of the virtual address into a corresponding physical address from a main memory; and store the translation of the virtual address into the corresponding physical address in a translation lookaside buffer (TLB) of the IOMMU.

Abstract: Machine learning (ML) and probabilistic techniques for identification and content delivery are described. A method comprises receiving, by a content delivery server, a request from a web server for content. The server then determines a set of user-probability pairs, each pair including a predicted user identifier and a probability that the predicted user identifier is associated with the request from the web server. The method further includes applying an objective function to the set of user-probability pairs to select one of the user identifiers with a corresponding candidate user profile. Based on the selected user identifier, a content item is selected for delivery to the web server.

Abstract: The present disclosure relates to systems and methods for transmitting Standard Dynamic Range (SDR) content. The systems and methods may use a modified Electro-Optical Transfer Function (EOTF) curve to convert nonlinear color values of SDR content into optical output values of modified SDR content. The systems and methods may encode the modified SDR content using eight bits while preventing banding. The systems and methods may transmit the encoded data to a client device for presentation on a display.

Abstract: This document relates to utilizing spare cores to mitigate hardware failures on a computing device. For instance, in some cases, tenant workloads can have sensitive data that executes on a given core of a device. In some cases, a hardware fault in the system may occur that affects the core running the tenant workload, and there may be an action that can be taken in mitigation software or firmware to fix that hardware fault. However, because of the sensitivity of the data on the core experiencing the fault, it is not always possible to run the mitigation algorithms on the core that is affected by the fault. The disclosed implementations can utilize a spare core to execute a hardware mitigation algorithm on another core that can clear the fault, without accessing any sensitive data on the core that is affected by the fault.

Abstract: A method for exciting a resonator at an estimated resonance frequency of the resonator comprises: (a) varying a frequency of an input signal applied to the resonator while recording an output signal from the resonator; (b) for each of a series of frequency values of the input signal, computing noise-filtered, complex first and second derivatives of a ratio of the output signal to the input signal; (c) for each of the series of frequency values of the input signal, computing a real part of a dot product of the complex first and second derivatives as vectors projected onto a complex plane; and (d) exciting the resonator an excitation frequency where the real part vanishes.

Abstract: Data may contain personal information and be subject to privacy requirements. The data may be encrypted and only a secure enclave may be able to decrypt the encrypted data. The secure enclave may be used to generate a report based on the encrypted data and a first set of added noise. The report may be subject to audit requirements and satisfy a differential privacy guarantee. The encrypted data may be stored for a first period. After the first period, the secure enclave may be used to generate a private synopsis based on the encrypted data and a second set of added noise. The private synopsis may satisfy the differential privacy guarantee. The private synopsis may be encrypted and only the secure enclave may be able to decrypt the encrypted private synopsis. The encrypted data may be expunged, and the encrypted private synopsis may be retained for a second period.

Abstract: A system comprises a substrate having a planar surface; a first magnet configured to apply a first magnetic field parallel to the planar surface; a circuit arranged on the planar surface; and a superconducting quantum interference device, SQUID, operably linked to the circuit. The SQUID comprises a Josephson junction arranged in a superconductive loop. The superconductive loop includes a portion which extends perpendicular to the planar surface and is orientated such that the SQUID is tuneable by the first magnet. By allowing the SQUID to be tuned using a magnetic field which is parallel to the planar surface, a reduction in flux noise may be achieved. Also provided are a method of operating a SQUID, and a SQUID.

Abstract: Example implementations include a method, apparatus, and computer-readable medium configured for generating codelets. An interface between a user and a large language model receives a natural language intent for a codelet to be executed within a network function to output one or more requested pieces of information from the network function. A codelet generation application generates a prompt to the large language model to write a codelet in an imperative procedural language to output the one or more requested pieces of information, the prompt including a hook point of the network function and one or more data fields of the network function associated with the one or more requested pieces of information. The application receives, in response to the prompt, imperative procedural language code for the codelet. The application statically verifies whether the procedural language code for the codelet satisfies constraints for execution within the network function.

Abstract: Systems and methods for state clearing of internal memory and flip-flops of a system using built-in test and scan circuitry are provided. A method includes a security processor: (1) fetching memory built-in self-test (MBIST) data and providing the MBIST data to an internal field-test (IFT) controller such that the IFT controller can drive the MBIST data via an IJTAG network to respective MBIST controllers, and (2) fetching scan data and providing the scan data to the IFT controller such that the IFT controller can drive the scan data via a scan network to respective embedded deterministic test (EDT) controllers. The method further includes each of the respective MBIST controllers state clearing respective internal memories by writing pertinent MBIST data into the respective internal memories. The method further includes each of the respective EDT controllers state clearing respective flip-flops by writing pertinent scan data into the respective flip-flops.

Abstract: A method of writing data to a transparent substrate comprises forming a first voxel by focusing a first laser pulse on a first location in a transparent substrate; and forming a second voxel by focusing a second laser pulse on a second location in the transparent substrate. The first laser pulse and the second laser pulse have different amplitudes, resulting in the first and second voxels having different strengths. Also provided are a system useful for implementing the method; an optical data storage medium obtainable by the method; and a method of reading data from the optical data storage medium.

Abstract: In a cloud computing environment, a cross-tenant access security measure monitors conditional access policies for changes or additions that hamper or threaten an authorized access from an assistant tenant user to a focus tenant. Some cross-tenant access security tracks role assignments to detect rogue roles, or detect hampering role changes. In some cases, focus tenant events and assistant tenant events are correlated in an audit. In some cases, the authorized access is a zero standing time bound access. In some cases, the authorized access is constrained to an IP address range, or constrained to login from a managed device, or both. In some cases, assets are excluded from managed response remediation actions. In some, managed response is modulated by product-specific Role Based Access Control. In some, repeated logins are avoided, to permit faster managed responses.

Abstract: Methods, systems, and computer storage media for providing unified device management using a unified device management engine of a device management system are described. Device management refers to administering, controlling, and maintaining various types of computing devices. The unified device management engine provides a unified policy object that supports unified management of a remote client and a local client using a unified policy. A unified policy object for a remote client and a local client is configured. The unified policy object is a combined management configuration that enables enforcement of a unified policy on both the remote client and the local client. The unified policy is communicated to the remote client. Unified device management data associated with the unified policy, the remote client, and the local client is generated. The unified device management data is communicated to cause display of the unified device management data on a device management interface.

Abstract: The present disclosure provides methods, systems and computer readable media for training and implementing a generative machine learning model for identifying and mitigating security threats. Certain examples relate to generative model training, in which a training image is provided to a generative machine learning (ML) model in a training prompt, with an Indicator of Compromise (IoC) prediction instruction pertaining to the first security image. The model generates a predicted IoC and a parameter of the model is updated based on a loss function that quantifies error between a ground truth IoC and the predicted IoC. Other examples relate to the use of trained generative models for cybersecurity. A mitigation prompt comprising a second security image and an associated mitigation instruction is provided to a trained generative model. The model outputs an indication of a cybersecurity mitigation action based on the mitigation prompt, and the cybersecurity mitigation action is performed on the system.

Abstract: The performance of a text parser implemented with a state machine is improved by reducing a critical dependence path. In one aspect, all possible current states for a given text input are read from a state table circuit, and the correct next state and output are then selected therefrom by an output multiplexer based on the current state, removing dependence on the current state from the table read, and allowing the read(s) to be pipelined. Further, multiple input units are configured to operate on multiple text characters in parallel, with each input unit propagating outputs for its state table circuit to the next downstream input unit. Each downstream input unit is configured to use the propagated states to provide the proper outputs to appropriates multiplexer inputs. The number of possible output states may be dynamically reduced, thereby reducing the size of the output multiplexer needed to select the next state.

Abstract: A computer-implemented method comprising detecting, by an application in a focussed operating mode, a trigger event; determining, using machine learning (ML) content classification applied to content associated with the trigger event, that content associated with the trigger event does not match a permitted topic associated with the focussed operating mode; and based on determining that the second content does not match the permitted topic applicable to the focussed operating mode, suppressing, by the application in the focussed operating mode, notification of the second trigger event.

Abstract: Systems and methods are provided for improving a reflection signal of a radar antenna assembly. For example, disclosed embodiments are directed to a multiple retrodirective surface mount antenna assembly comprising a plurality of retrodirective antenna arrays, each retrodirective antenna array being angularly offset from at least one other retrodirective antenna array and being configured to reflect a radar signal received from a radar signal source to a radar signal receiver.

Abstract: A computer system performs operations to prepare input to a generative artificial intelligence (“AI”) model. The system receives compressed data for media, which has been compressed according to a media compression format to produce the compressed data. The system partially decompresses the compressed data (e.g., performing parsing and entropy decoding operations). This produces syntax elements of the compressed data according to the media compression format. The system converts the syntax elements into tokens that represent the syntax elements, respectively. Unlike the syntax elements (in the media compression format), the tokens are encoded in an input format for the generative AI model. The system stores the tokens in memory or storage, from which the system can provide the tokens to the generative AI model for use in a training process or inference process for media synthesis, media compression, media decompression, or another purpose.

Abstract: Techniques are described for decoding frame masks in a video stream. In some solutions, a computer implemented method receives a frame of video data comprising an array of pixels with encoded color values defined in a YUV color space. The encoded values, converted from original color values using a mapping function, are bounded within a second data range, shifted or compressed from a first data range of the original color values. The method decodes the encoded color values to obtain decoded color values for the array of pixels and extracts a foreground mask based on the decoded color values. The foreground mask identifies a subset of the array of pixels to be in a foreground of the frame. The method then converts the decoded color values of the subset of the array of pixels to new color values by applying an inverse mapping function.

Abstract: Methods, systems, apparatuses, and computer-readable storage mediums described herein enable executable code of a hardware security platform (HSP) circuit to communicate with a hypervisor in a separate processor. The hypervisor generates and manages virtual machines. The HSP code comprises trusted platform module (TPM) logic, that processes TPM commands received via the hypervisor, and in response to the processing, communicates security information (e.g., measurements, keys, authorization data) with the virtual machines via the hypervisor. The TPM logic receives security information related to a virtual machine from the hypervisor and stores the security information in non-volatile memory of the HSP circuit, where security information from a particular VM is distinguishable from security information from another VM in the HSP memory.

Abstract: A method, implemented at a computer system in a network, for localizing network failure in the network includes enumerating at least one path between the computer system and a path destination node in the network through one or more intermediary nodes. The method further includes enumerating a plurality of sub-paths between the computer system and the path destination node. Each sub-path has either the path destination node, or one or more intermediary nodes of the corresponding path as a corresponding sub-path destination node. For each of the plurality of sub-paths, determining a probability that a data packet can successfully traverse the round trip of the corresponding sub-path and identifying a node in the network where network failure have occurred.

Abstract: Systems are configured for generating and utilizing training data to train learn-to-rank type models in a manner that preserves privacy of client data used for generating the training data. The systems extract features and patterns of the user queries, search results and user interactions with the search results without tracking, storing or transmitting underlying values of the user data to preserve privacy of the user data. Systems are also configured to infer search result quality based on at least the user behavior data, and optionally query intentions, and to generate and label corresponding training data accordingly. This training data is applied to learn-to-rank type models to train the learn-to-rank type model to improve search quality of search results provided by the learn-to-rank type models when new user queries are processed that having features and patterns corresponding to the filtered and labelled training data.

Abstract: High availability network services are provided in a communications network comprising a plurality of network devices including a network function implemented as two instances configured as an active instance and a backup instance. The backup instance maintains state data such that the backup instance can actively provide services in response to a failure of the active instance. A pool of data forwarding functions sends, over a tunnel connection, ingress data packets to the network function based on a MAC address of the active instance on an overlay network. When the active instance has failed, the backup instance provides the network function and the pool of data forwarding functions sends over the tunnel connection, subsequent ingress data packets to the network function based on an overlay network MAC address of the backup instance.

Abstract: Systems and methods are disclosed for credit-based flow control of multiple data links using a common reverse channel. The links transfer data from a source to respective buffers at a sink. Credits represent available buffer space. For each data link, a credit counter at the source is decremented as data is transmitted and incremented as the sink returns credits. Reporting logic at the data sink generates a credit report as sink logic retrieves data from the buffer, freeing buffer space. Encoding logic aggregates the credit reports from multiple links for transmission over the common reverse channel to the source, where individual credit reports are extracted and distributed among the links, for update to the respective credit counters. For each link, data transmission pauses when the credit counter decreases to a threshold. Return multiple links' credits over a single reverse channel saves power. Variations are disclosed.

Abstract: Examples are disclosed relating to an approach for individualized image compression. In one example, an individual contrast sensitivity function (CSF) of a user is received from a remote computing system. A distance measurement indicating a distance between the user and a display of the remote computing system is received from the remote computing system. An image frame is generated. One or more parameters of a compression algorithm are modified based at least on the individual CSF of the user and the distance measurement. A compressed image frame is generated from the image frame using the compression algorithm based at least on the one or more modified parameters.

Abstract: Digital image processing methods performed by a computer are disclosed. In one example, a first digital image captured by a real camera is sub-divided into a first plurality of sub-images. A second digital image captured by a real camera is sub-divided into a second plurality of sub-images. A set of image features in a first sub-image of the first plurality of sub-images is identified. A subset of neighboring sub-images is identified from the second plurality of sub-images based at least on each neighboring sub-image of the subset of neighboring sub-images having one or more corresponding image features in common with the set of image features identified in the first sub-image.

Abstract: A method of AI content evaluation includes receiving, from a generative artificial intelligence (AI) model, a set of AI-generated instructions that identifies steps for performing a task within an application, and selecting checkpoint interactions from an interaction index that define a plurality of interactions with a user interface. Each of the checkpoint interactions satisfies a similarity metric with a corresponding step in the set of AI-generated instructions. The method further includes determining, based on detected user interactions with the user interface, a subset of the checkpoint interactions completed by a user within an observation period, and evaluating a metric that to compute a quality score that quantifies user success with respect to performing the task associated with the AI-generated instructions. The metric depending at least in part on the subset of the checkpoint interactions completed by the user within the observation period.

Abstract: A method for utilizing contextual data in generating machine translations. The method includes receiving a translation request including an initial prompt received via a user interface. The initial prompt includes a first language passage and a translation instruction. The initial prompt also includes a context data signal received via a context data source. The method further includes generating a context instruction based on the context data signal and generating a modified prompt including the initial prompt and the context instruction. The method further includes sending the modified prompt to a neural machine translation (NMT) model to process the modified prompt and receiving a second language translation passage as a response to the modified prompt. The second translation language passage being a second language translation of the first language passage translated according to the translation instruction and the context instruction.

Abstract: Branch target buffer structures are provided. A device can include a hierarchy of branch target buffers storing entries corresponding to branch instructions, the hierarchy of branch target buffers including respective branch target buffers that have progressively slower access times. The device can include a first program counter configured to generate a first program counter value associated with a next instruction of an executing application. The device can include a second program counter configured to predict a second program counter value that is associated with a subsequent instruction of the executing application that is after the next instruction. The device can include first branch prediction circuitry configured to populate a branch target buffer of the branch target buffers based on the second program counter value.

Abstract: Aspects of the present disclosure relate to semantic user input for a computing device. In examples, user input is identified and processed to identify and automatically perform an associated semantic action. The semantic action may be determined based at least in part on an environmental context associated with the user input. Thus, an action determined for a given user input may change according to the environmental context in which the input was received. For example, an association between user input, an environmental context, and an action may be used to affect the behavior of a computing device as a result of identifying the user input in a scenario that has the environmental context. Such associations may be dynamically determined as a result of user interactions associated with manually provided input, for example to create, update, and/or remove semantic actions associated with a variety of user inputs.

Abstract: The disclosure provides a method and an apparatus for song generation. A text input may be received. A topic and an emotion may be extracted from the text input. A melody may be determined according to the topic and the emotion. Lyrics may be generated according to the melody and the text input. A song may be generated at least according to the melody and the lyrics.

Abstract: Examples are disclosed that relate to an augmented-reality device including a near-eye display that is configured to provide display uniformity at varying operating temperatures. The near-eye display includes a laser light source that is configured to perform a display scan operation including, for each pixel of a plurality of pixels that collectively form a field of view of the near-eye display, generating a pre-color pulse and subsequently generating a primary color pulse. The augmented-reality device is configured to receive an operating temperature measurement of the laser light source from a temperature sensor. The augmented-reality device is further configured to, for each pixel of the plurality of pixels, dynamically tune an amplitude of the pre-color pulse based at least on the operating temperature measurement of the laser light source, and generate, via the laser light source, the pre-color pulse with the dynamically tuned amplitude, and subsequently generate the primary color pulse.

Abstract: Phased loss function training of machine learning models that are configured for performing super-resolution image processing includes training with two or more discrete phases that each apply different combinations of loss function training to the super-resolution models. In a first phase, for example, a first loss function is applied to the model that is a per-pixel or a non-perceptual loss function. In a second phase, which begins after the model reaches a threshold of convergence in the first phase, a blended loss function is applied, which includes the application of at least one perceptual loss function. Then, after the model reaches a subsequent threshold of convergence in the second phase, which is greater than the threshold of convergence in the first phase, the model is optionally further modified with optimizations such as quantization and/or sparsity optimizations for further facilitating subsequent super-resolution processing.

Abstract: A system for selecting motion models for aligning scene content captured by different image sensors, is configurable to (i) access a first image captured by a first image sensor and a second image captured by a second image sensor; (ii) access a set of motion models; (iii) define a reference patch within the second image; (iv) generate a respective match patch for each motion model of the set of motion models; (v) determine a similarity between each respective match patch and the reference patch within the second image; (vi) select a final motion model from the set of motion models based upon the similarity between each respective match patch and the reference patch within the second image; and (vii) utilize the final motion model to generate an output image for display to a user.

Abstract: The techniques disclosed herein enable systems to offload deep packet inspection tasks to a network interface card. This is accomplished by configuring the network interface card with a configuration file. The configuration file identifies target protocols, target fields, a number of packets to analyze for each target protocol, as well as identification tables that enable the network interface card to identify packet attributes. Once configured, the network interface card can receive and analyze incoming network packets. Accordingly, the network interface card extracts and parses values represented by the network packet in accordance with the parameters of the configuration file. The extracted values are compared against the entries of the identification table to derive an attribute identifier which can be returned to a network protocol stack. Moreover, the configuration file can also provide support for standard and non-standard network protocols.

Abstract: An electronic device includes a device housing and a secondary component removably affixed to the device housing. A dual screw assembly is used to affix the secondary component to the device housing. The dual screw assembly includes a first screw inserted into a housing aperture of the device housing, such that a portion of the first screw extends into an interior of the device housing, and a second screw inserted into a screw aperture formed within a head of the first screw, such that at least a portion of the secondary component is positioned between a head of the second screw and the head of the first screw.

Abstract: A mixed-reality waveguide combiner for use in an optical display system of a head-mounted display (HMD) device for mixed-reality applications in which virtual images are displayed over views of the real world by an HMD device user is disclosed. A gradient refractive index is utilized in a surface relief grating having depth-modulated grating structures along with a thickness-modulated residual layer. A foundation layer, providing a contrasting refractive index to that of the gratings and the residual layer, may be utilized to provide increased resistance against parasitic diffraction away from the HMD device user towards the surrounding world-side environment. Such world-side leakage causes a phenomenon known as “eye glow” which can reduce social comfort by obscuring the user's eyes or undesirably increase HMD device observability.

Abstract: A technique processes input information using a machine-trained model in a context-specific manner. The technique includes, for a particular processing block: receiving a set of input tokens; transforming the set of input tokens into a set of output tokens; and generating a control decision based on the output tokens. For a first state, the technique appends at least one of the output tokens to the input tokens to define a new set of input tokens and then repeats the transforming and generating operations. For a second state, the technique passes at least some of the set of output tokens to a next processing block. The token(s) passed to the next processing block convey supplemental information that does not necessarily have a verbal counterpart. The technique has the end effect of invoking transformation operations in a context-based as-needed basis, which, in turn, reduces latency and the consumption of resources.

Abstract: A data processing system implements receiving a text prompt to create a multilayer graphic design; predicting, by a first generative model based on the text prompt, a layout including anonymous regions each defined by a bounding box without content or region-wise prompt annotations; concurrently generating, by a diffusion transformer, multilayer image latents of a global reference image, a background layer, and transparent foreground layers using a Gaussian noise conditioned on the layout and the text prompt, each of the transparent foreground layers corresponding to one of the anonymous regions; decoding, by a vision transformer, the multilayer image latents into the global reference image, the background layer, and the transparent foreground layers as the multilayer graphic design; composing an output based on the background layer and the transparent foreground layers; providing the output to a client device; and causing a user interface of the client device to display the output.

Abstract: One or more digital agents are instantiated to monitor to monitor one or more communication sessions on behalf of a user. Respective digital agents are configured to identify, in the information shared during respective communication sessions, respective one or more subsets of information including one or more items of information relevant to the user. The respective one or more subsets of information identified, by the respective one or more digital agents, are obtained from the one or more digital agents. At least some of the information in the respective subsets of information is presented to the user.

Abstract: In various examples there is a computer-implemented method of generating a patch to fix a vulnerability in a downstream software package. An upstream source code and an upstream patch are received. A downstream software package corresponding to the upstream source code is identified. A training shot is selected from a store of training shots. Input for a generative model is generated, the input comprising the training shot and an inference shot, the inference shot comprising the upstream source code, upstream patch, rejected hunks of the upstream patch and downstream source code. The input is provided to the generative model which generates the downstream patch.

Abstract: A speculative decoding technique generates tokens of a response using a verifying model and drafting model. In operation, the verifying model receives a set of draft tokens produced by the drafting model. The verifying model then verifies whether the draft tokens are correct by comparing probability information generated by the verifying model with probability information generated by the drafting model. The verifying model then determines whether any draft token that is rejected by the verifying is otherwise accepted based on a specified matching criterion. The above process is repeated one or more times to generate the response. The technique has the overall effect of expanding a number of draft tokens that are accepted, thereby reducing use of resources and deceasing latency.

Abstract: A computing system including one or more processing devices configured to receive an input query. At a query parsing machine learning (ML) model, the one or more processing devices extract a temporal substring from the input query. At a sentence embedding ML model, the one or more processing devices compute a sentence embedding of the temporal substring. At a date classifier ML model, the one or more processing devices compute a date embedding based at least in part on the sentence embedding and perform classification on the date embedding to identify an output date including a day, a month, and a year indicated by the temporal substring. The one or more processing devices output the output date to an additional computing process.

Abstract: A parallel processing drawing method operable on a computing device, such as a mixed- or virtual-reality head-mounted display (HMD) device, is utilized to provide a drawing tool that enables a user to virtually draw a line on a mesh of primitive geometric objects (e.g., triangles) that model three-dimensional (3D) surfaces of virtual or real-world objects. A device user provides inputs to a user interface on the computing device to virtually draw a line that the drawing tool then renders on a display system. The parallel processing drawing method simultaneously computes each of the individual line segments to conform the virtually drawn line to the mesh. The method is computationally efficient compared to conventional ray tracing which facilitates applications such as quick marking and drawing on meshes including complex meshes representing terrain and irregularly shaped objects.

Abstract: A machine reading comprehension system (MRCS) can analyze a larger-sized document that includes multiple pages to predict an answer to a query. For example, the document can have two, five, tens, or hundreds of pages. The MRCS divides the document into multiple sections with each section including a portion of the document. Each section is processed separately by one or more processing circuitries to determine a score for that section. The score indicates how related the section is to the query and/or a probability that the section provides a possible answer to the query. Once all of the sections have been analyzed, the sections are ranked by their scores and a subset of the ranked sections are processed again to determine a predicted answer to the query.