A training system trains a floor plan-encoding machine-trained model (“model”) that converts an image that depicts a floor plan into a universal latent-space floor plan representation. The floor plan representation is referred to as universal or uniform because different floor plans that depict the same building layout in different ways map to the same latent-space floor plan representation. The training system also trains an image-encoding model that maps a camera image that depicts a part of a building layout to a latent-space image representation. A fusing operation integrates the latent-space image representation into a spatially-matching part of a latent-space floor plan representation produced by the floor plan-encoding model. The approach used by the training system eliminates the need to train and maintain different downstream models.
G06V 10/774 - Génération d'ensembles de motifs de formationTraitement des caractéristiques d’images ou de vidéos dans les espaces de caractéristiquesDispositions pour la reconnaissance ou la compréhension d’images ou de vidéos utilisant la reconnaissance de formes ou l’apprentissage automatique utilisant l’intégration et la réduction de données, p. ex. analyse en composantes principales [PCA] ou analyse en composantes indépendantes [ ICA] ou cartes auto-organisatrices [SOM]Séparation aveugle de source méthodes de Bootstrap, p. ex. "bagging” ou “boosting”
G06V 10/82 - Dispositions pour la reconnaissance ou la compréhension d’images ou de vidéos utilisant la reconnaissance de formes ou l’apprentissage automatique utilisant les réseaux neuronaux
G06V 20/58 - Reconnaissance d’objets en mouvement ou d’obstacles, p. ex. véhicules ou piétonsReconnaissance des objets de la circulation, p. ex. signalisation routière, feux de signalisation ou routes
2.
NEAR REAL-TIME ARTIFICIAL INTELLIGENCE (AI) ENHANCEMENT OF VIDEO OR AUGMENTED REALITY (AR) PRESENTATIONS OF LIVE EVENTS
Systems and methods are provided for implementing near-real-time artificial intelligence (“AI”) enhancement of video or augmented reality (“AR”) presentations of live events. In examples, a computing system identifies, from a plurality of spoken words in a live audio recording of individuals speaking, spoken words to modify, based on a set of parameters and/or a profile of a user, using an AI agent. The computing system causes a generative AI system to modify, in near-real-time, (a) the identified spoken words with corresponding replacement words in each of the live video recording and the corresponding live audio recording, based on the profile of the user; (b) portions of the live video recording and the corresponding live audio recording to generate video and audio smooth transitions before and after the replacement words are incorporated in the live video recording and the corresponding live audio recording, respectively.
G06F 40/58 - Utilisation de traduction automatisée, p. ex. pour recherches multilingues, pour fournir aux dispositifs clients une traduction effectuée par le serveur ou pour la traduction en temps réel
G10L 15/10 - Classement ou recherche de la parole utilisant des mesures de distance ou de distorsion entre la parole inconnue et les gabarits de référence
G10L 15/30 - Reconnaissance distribuée, p. ex. dans les systèmes client-serveur, pour les applications en téléphonie mobile ou réseaux
G10L 25/57 - Techniques d'analyse de la parole ou de la voix qui ne se limitent pas à un seul des groupes spécialement adaptées pour un usage particulier pour comparaison ou différentiation pour le traitement des signaux vidéo
The techniques disclosed herein are directed to generating recap summaries for an ongoing event (e.g., an online meeting, a podcast) that are customized for late-joining users. As mentioned above, generative artificial intelligence-based productivity tools are often used for summarizing audiovisual content such a text document, an online meeting, and the like. However, many existing artificial intelligence systems lack the ability to adapt summary outputs to suit the needs of individual users leading to a frustrating user experience. In contrast, the present system and method provides mechanisms for customizing the retrieval of audiovisual content corresponding to an elapsed portion of an ongoing event based on the time of entry of a late-joining user thereby enabling adaptive recap summaries. Moreover, the content of the summaries can be further customized based on contextual information such as a user’s role (e.g., a job title) thereby ensuring helpful recap summaries.
H04L 12/18 - Dispositions pour la fourniture de services particuliers aux abonnés pour la diffusion ou les conférences
H04L 51/063 - Adaptation du contenu, p. ex. remplacement d'un contenu inapproprié
H04L 65/613 - Diffusion en flux de paquets multimédias pour la prise en charge des services de diffusion par flux unidirectionnel, p. ex. radio sur Internet pour la commande de la source par la destination
4.
Compressing Information Provided to a Machine-Trained Generative Model
A technique is described for compressing input information fed to a machine-trained generative model. The technique includes: receiving original input information having a plurality of sentences; performing word-level encoding of the original input information using a first part of a machine-trained transformer model, to provide word-level encoded information; performing sentence-level encoding of the word-level encoded information using a second part of the machine-trained transformer model, to provide scores associated with the first plurality of sentences; selecting a subset of the sentences in the original input information based on the scores, to provide modified input information; and providing the modified input information to the machine-trained generative model. The operation of word-level encoding performs parallel processing on portions of the original input information. Overall, the compressing operation enables the machine-trained generative model to process the original input information with reduced resources and latency.
Systems and methods are provided for implementing post-production artificial intelligence (“AI”) enhancement of video presentations based on script modification. In examples, a computing system identifies, from a plurality of spoken words in an audio file of individuals speaking, spoken words to modify, based on a set of parameters and/or a profile of an intended audience, using an AI agent. If there is a script for the audio file, the computing system causes a generative AI system to modify the script to replace the identified spoken words with replacement words, and to modify corresponding portions of the audio file to replace the identified spoken words with the replacement words based on the modified script. The computing system further causes the generative AI system to generate audio transitions before and after the replacement words, and to modify the applicable portions of a corresponding video file based on the modified audio file.
G06F 40/166 - Édition, p. ex. insertion ou suppression
G10L 15/183 - Classement ou recherche de la parole utilisant une modélisation du langage naturel selon les contextes, p. ex. modèles de langage
G10L 25/57 - Techniques d'analyse de la parole ou de la voix qui ne se limitent pas à un seul des groupes spécialement adaptées pour un usage particulier pour comparaison ou différentiation pour le traitement des signaux vidéo
A cryptographic device performs a method including installing a device firmware update in the cryptographic device that includes a device identifier composition engine (DICE). In response to installing the device firmware update and during a boot operation of the device, performing, using a unique device secret (UDS), a first set of cryptographic operations to generate a first set of deterministic signatures. A tuned nonce is selected based on a selected number of signature generation rounds. A non-deterministic signature is generated based on the tuned nonce during the boot operation. The boot operation is completed to enable run time generation of signatures.
Techniques are described herein that are capable of performing over-subscription ratio-based sharding in an AI system. The AI system includes a first rack of first AI accelerators, NICs associated with the first AI accelerators, and a second rack of second AI accelerators. Subsets of the first AI accelerators establish a consensus regarding allocation of a workload among the first AI accelerators by communicating via the NICs. The consensus indicates portions of data that are to be transmitted from the first rack to the second rack via a Tier1 switch and time durations to be consumed by transmission of the portions. The first AI accelerators group the portions into shards corresponding to respective timeslots based at least on the time durations and an over-subscription ratio associated with the first AI accelerators. Designated first AI accelerators transmit the shards in the respective timeslots to the second rack via the Tier1 switch.
The disclosure is directed to systems, methods, and computer storage media, for, among other things, employing nested model structures to enforce compliance, within a computational system, to at least one policy. One method includes receiving a digital record that encodes content. A plurality of models (e.g., integrated models and/or model droplets) is employed to analyze the records. The plurality of models is configured and arranged within a nested structure of a hierarchy of models. Each of the plurality of models analyzes at least a portion of the record. Based on the nested structure, the hierarchy combines the analysis from each of the plurality of models to determine that the content violates a policy of a system. In response to determining that the content violates the policy, at least one mitigation (or intervention) action are performed. The at least one mitigation action may alter subsequent transmissions of the record.
G06Q 10/0637 - Gestion ou analyse stratégiques, p. ex. définition d’un objectif ou d’une cible pour une organisationPlanification des actions en fonction des objectifsAnalyse ou évaluation de l’efficacité des objectifs
G06F 16/28 - Bases de données caractérisées par leurs modèles, p. ex. des modèles relationnels ou objet
This document relates to automated classification of behavior structured differences in release engineering workflows. A computer-implemented method includes receiving a set of behavior structured differences generated from structured differential testing, where each behavior structured difference represents a discrepancy between outputs of a test environment and a production environment for the same real-world input. The method further includes extracting a plurality of attributes from each behavior structured difference, including at least test logs and production logs, and generating an input prompt for a language model (LLM) by structuring the attributes into a textual sequence. The LLM, trained on historical behavior structured differences labeled as noise, new feature, or regression, classifies each behavior structured difference into one of the categories based on the input prompt. The classification is then output to a user interface to support validation of a software update.
A method of manufacturing a cryptographically enabled device includes iteratively selecting an initial unique device secret (UDS) seed for a cryptographically enabled device, performing cryptographic signature generation based on the initial UDS seed, tracking performance of the cryptographic signature generation, and comparing performance of the cryptographic signature generation to a performance threshold until comparing identifies an optimal UDS seed. Once the optimal seed is identified in the above iteration, the optimal UDS seed is fused into the cryptographically enabled device.
H04L 9/32 - Dispositions pour les communications secrètes ou protégéesProtocoles réseaux de sécurité comprenant des moyens pour vérifier l'identité ou l'autorisation d'un utilisateur du système
A domain-specific agent may receive a multi-modal data signal associated with a reported performance issue and instructs a domain-specific multi-modal language model to identify a possible root cause for the reported performance issue. The domain-specific agent may retrieve, using the domain-specific agent, from a pre-populated hypothesis tree, a hypothesis that identifies the possible root cause and sets forth a series of operations for testing whether the possible root cause is linked to the reported performance issue. The domain-specific agent may instruct the domain-specific multi-modal language model to execute the series of operations for testing the hypothesis. In response to receiving an output from the domain-specific multi-modal language model identifying the hypothesis as true, the domain-specific agent generates and returns an output that identifies the possible root cause as a source of the reported performance issue.
A match table is used to match a concatenation of a source address of a data packet and a consecutive message sequence number (MSN) to a match data structure. The matched concatenation is inserted into the match data structure based on a source and a running counter for a MSN to each source. The currently active PDC associated with the source is attached or an empty PDC is atomically created and attached. In response to a packet arriving before the semantic layer has posted a recv(), A canary value “failed match” is atomically inserted into the match data structure and a request to wait (RTW) message is sent to the source. In response to the semantic layer posting the recv(), the “failed match” entry in the match data structure is identified, the “failed match” entry is atomically updated with match information, and a request to send (RTS) message is sent to the source.
H04L 47/12 - Prévention de la congestionRécupération de la congestion
H04L 47/24 - Trafic caractérisé par des attributs spécifiques, p. ex. la priorité ou QoS
H04L 47/263 - Modification du taux à la source après avoir reçu des retours
H04L 47/27 - Évaluation ou mise à jour de la taille de la fenêtre, p. ex. en utilisant des informations dérivées de paquets [ACK] d’acquittements
H04L 47/30 - Commande de fluxCommande de la congestion en combinaison avec des informations sur l'occupation de mémoires tampon à chaque extrémité ou aux nœuds de transit
H04L 47/33 - Commande de fluxCommande de la congestion en utilisant le transfert de la notification
H04L 47/43 - Assemblage ou désassemblage de paquets, p. ex. par segmentation et réassemblage [SAR]
H04L 47/62 - Ordonnancement des files d’attente caractérisé par des critères d’ordonnancement
13.
PERSISTENT SESSION LOAD BALANCING AND VIRTUAL MACHINE DEALLOCATION
Techniques are disclosed for dynamically distributing received requests for persistent sessions to virtual machines, and for selectively deallocating the virtual machines. A machine readiness status of a first virtual machine is determined. Received persistent session requests are routed to the first virtual machine when the determined machine readiness status has a first value. When the determined machine readiness status has a second value, subsequently received persistent session requests are routed to a second virtual machine, and no additional received persistent session requests are routed to the first virtual machine. The first virtual machine hosts persistent sessions corresponding to the received requests for persistent sessions routed to the first virtual machine. End times are determined for each hosted persistent session, and a maximum of the end times is determined. The first virtual machine is selectively deallocated based on the determined maximum end time.
H04L 67/1027 - Maintien des sessions pendant la répartition de la charge
G06F 9/455 - ÉmulationInterprétationSimulation de logiciel, p. ex. virtualisation ou émulation des moteurs d’exécution d’applications ou de systèmes d’exploitation
H04L 67/63 - Ordonnancement ou organisation du service des demandes d'application, p. ex. demandes de transmission de données d'application en utilisant l'analyse et l'optimisation des ressources réseau requises en acheminant une demande de service en fonction du contenu ou du contexte de la demande
14.
AUTOMATED INFORMATION CHECKING IN SYNTHESIZED TEXT DOCUMENT SUMMARIES
The techniques discussed herein are directed to real-time information checking when generating summaries of text documents. Generative artificial intelligence productivity tools are often used for summarizing content such a long text documents (e.g., research papers, legal filings, meeting transcripts). However, many existing artificial intelligence systems lack mechanisms for ensuring the content of the summaries they generate is accurate and/or complete. Consequently, this lack of functionality can lead to the propagation of inaccurate and/or misleading information as well as a disjointed, confusing user experience. To address this technical challenge, the present system and method provides mechanisms that enable generative artificial intelligence models to check the accuracy and/or completeness of information within text documents using a semantic comparison against information from trusted sources such as research institution databases. Moreover, the system provides functionality for seamlessly incorporating corrections into summaries to inform users of potentially inaccurate and/or incomplete information.
A computing system is provided, including processing circuitry configured to, in a training phase, receive a plurality of example multi-turn conversations between users and generative language models involving tasks, where each task has associated levels and/or types of empathy. The processing circuitry is further configured to generate synthetic multi-turn conversations for each task cluster based on empathy patterns, using a generative language model system, and train an expert adapter based on the synthetic multi-turn conversations to adapt an output of the generative language model system to achieve a task-specific target level and/or type of empathy. In an inference phase, the processing circuitry is further configured to receive an inference-time user query related to an inference-time task at the generative language model system, which includes a generative language model and the trained expert adapter, and generate a response to the inference-time user query using the generative language model and expert adapter.
Systems, methods, and computer program products are disclosed for automatic code generation and testing. Metadata is determined for a dataset. Data processing code for processing data in the dataset is automatically generated based on the metadata. Test cases are automatically generated for testing the data processing code based on the first data processing code and a data sample generated based on the dataset. The data processing code is automatically executed based on the test cases to generate actual test results. The data processing code is validated by comparing the actual test results to expected test results associated with the test cases.
The described technology provides a method including determining a cacheline sector associated with a snoop filter (SFT) having a plurality of SFT entries, determining a number of cachelines in the identified cacheline sector that are cached by one or more agents, and determining, based on the number of cachelines in the identified cacheline sector that are cached by one or more agents, a number of bits in a bit vector (BV) of one or more of the plurality of SFT entries, wherein the number of bits is required to track the one or more agents.
A method of performing a collective operation using electromagnetic energy includes displaying a first encoded electromagnetic signal representing a first value encoded via a property of electromagnetic energy as a first pattern of one or more entries and displaying a second encoded electromagnetic signal representing a second value encoded via the property of electromagnetic energy as a second pattern of one or more entries. The method includes detecting, at a set of regions of interest, the property of electromagnetic energy of a first entry of the first pattern and of the second pattern and determining a count for a first index position based on counting the properties of electromagnetic energy detected for the first entry of the first pattern and of the second pattern. The method includes determining a sum value of the first and second values based on the count and a multiplication factor for the first index position.
Systems and methods are provided for offloading a task from a central processor in a radio access network (RAN) server to one or more heterogeneous accelerators. For example, a task associated with one or more operational partitions (or a service application) associated with processing data traffic in the RAN is dynamically allocated for offloading from the central processor based on workload status information. One or more accelerators are dynamically allocated for executing the task, where the accelerators may be heterogeneous and my not comprise pre-programming for executing the task. The disclosed technology further enables generating specific application programs for execution on the respective heterogeneous accelerators based on a single set of program instructions. The methods automatically generate the specific application programs by identifying common functional blocks for processing data traffic and mapping the functional blocks to the single set of program instructions to generate code native to the respective accelerators.
The present disclosure generally relates to computer hardware, configurations, and methods for increasing the speed and efficiency in artificial intelligence (AI) accelerator racks with a fixed communication topology. For example, the computer hardware, configurations, and methods described herein replace the network switch typically included in an AI accelerator rack with crossbar switches that are added to trays of AI accelerators within the rack. One or more embodiments can include a tray of AI accelerators with both a y-dimension crossbar switch and an x-dimension crossbar switch. Both the y-dimension crossbar switch and the x-dimension crossbar switch may be configured to a fixed AI model communication topology that enable the AI accelerators within the tray to communicate rapidly across the y-dimension crossbar switch and the x-dimension crossbar switch with very little power expenditure. Additionally, an AI accelerator rack can include additional z-dimension crossbar switches across trays of AI accelerators.
A method of performing a collective operation using electromagnetic energy includes displaying a first encoded electromagnetic signal representing a first value encoded via a property of electromagnetic energy as a first pattern of one or more entries and displaying a second encoded electromagnetic signal representing a second value encoded via the property of electromagnetic energy as a second pattern of one or more entries. A first entry of the first pattern and a first entry of the second pattern are each displayed at a first shared entry window and are at least partially collocated. Based on detecting and counting the properties of electromagnetic energy at the first shared entry space, a count is determined for a first index position. The method includes determining a sum value of the first and second values based on the count and a multiplication factor for the first index position.
This disclosure presents a technique for privacy preserving capture of audio or other modality from an environment during a conversation that only records the voice of a user of an audio capture device. Raw audio captured from the environment is diarized into separate voice streams for each voice. The voice stream of the user of the audio capture device is identified. The raw audio and voice streams of conversation participants other than the user are discarded. A transcript is generated from the voice stream of the user. The transcript may be used for training a personalized artificial intelligence (AI) assistant or generating derived content. If permission is granted, voice streams of other participants or transcripts of their portions of conversation may be used to provide fuller context.
An action index includes information regarding an actionable graphical user interface (GUI) element of a GUI, an action for the actionable GUI element, a precondition for the action, and a postcondition for the action. A presence of the precondition in the GUI indicates a likelihood of success of the action, and a presence of the postcondition in the GUI, after performing the action, indicates success of the action. An agent performs a task on behalf of a user according to a plan that is based on the action index. The agent comprises an action engine which receives the task and obtains, from a generative artificial intelligence (GAI) model, the plan for completing the task.
G06F 3/0484 - Techniques d’interaction fondées sur les interfaces utilisateur graphiques [GUI] pour la commande de fonctions ou d’opérations spécifiques, p. ex. sélection ou transformation d’un objet, d’une image ou d’un élément de texte affiché, détermination d’une valeur de paramètre ou sélection d’une plage de valeurs
G06F 9/451 - Dispositions d’exécution pour interfaces utilisateur
Technology disclosed herein includes systems and methods for semantic text column augmentation in tabular datasets. In an embodiment, an augmentation system generates embeddings that represent semantic characteristics of textual entries in a tabular dataset and applies a selection algorithm to the embeddings to identify a diverse subset of the entries. The augmentation system sends the diverse subset to a language model to obtain labels for the diverse subset, receives the labels from the language model, and then determines semantic similarity between the embeddings of the diverse subset and the embeddings of the remaining textual entries. The system further assigns labels to the remaining textual entries based on the labels for the diverse subset and the semantic similarity between the embeddings.
A method, computer program product, and computing system for generating verified prompt-response pairs to reduce artificial intelligence (AI) model hallucinations include extracting a portion of content from a verified document, generating, using an artificial intelligence (AI) model, a prompt associated with subject matter included in the portion of content, generating, using the AI model, a plurality of responses to the prompt based on the prompt and the portion of content, selecting a response from the plurality of responses based on similarity to the portion of content, and storing the prompt and response as a verified prompt-response pair.
A method of fabricating a semiconductor device includes providing a workpiece (200). The workpiece comprises a semiconductor structure (210), the semiconductor structure comprising a substrate (212) and a semiconductor component (214) arranged on the substrate; a capping layer (220) over the semiconductor structure; and a shadow wall (230) comprising a leg and a body connected to the leg. The capping layer is a layer of arsenic or antimony. The leg of the shadow wall extends from the semiconductor structure through an aperture in the capping layer. The body of the shadow wall is on the capping layer and over the semiconductor structure. The method further comprises removing the capping layer, and, after removing the capping layer, depositing a further material onto a first region of the semiconductor structure. The shadow wall prevents deposition of the further material onto a second region of the semiconductor structure. Also provided is a workpiece useful in the method.
A method (400) for calibrated touch sensing on an electronic device (102) comprises: scanning (446B) driven electrodes (222) of a touch sensor (204) while recording signal from sensory electrodes (224) of the touch sensor; storing (446C) the signal in a signal matrix; for each element in the signal matrix that is lower than a predetermined threshold, adjusting (446F) a corresponding element in the calibration matrix based at least partly on a value of the element in the signal matrix and on a value of the corresponding element in the calibration matrix; for each element in the signal matrix that is above the predetermined threshold, estimating (446G) the corresponding element in the calibration matrix based at least partly on one or more elements neighboring the corresponding element in the calibration matrix; and returning (446H) a calibrated signal value based at least partly on corresponding values in the signal and calibration matrices.
A method for optical multiplication includes displaying a first encoded electromagnetic signal on a shared memory space, the first encoded electromagnetic signal being based on an input value and displaying a second encoded electromagnetic signal on the shared memory space, the second encoded electromagnetic signal being based on a learned parameter associated with the input value. The method includes detecting a cumulative electromagnetic signal on the shared memory space including a combination of at least some of the first encoded electromagnetic signal and at least some of the second encoded electromagnetic signal and determining a product of the input value and the learned parameter based on decoding the cumulative electromagnetic signal.
The techniques disclosed herein are directed to generating recap summaries for an ongoing event (e.g., an online meeting, a podcast) that are customized for late-joining users. As mentioned above, generative artificial intelligence-based productivity tools are often used for summarizing audiovisual content such a text document, an online meeting, and the like. However, many existing artificial intelligence systems lack the ability to adapt summary outputs to suit the needs of individual users leading to a frustrating user experience. In contrast, the present system and method provides mechanisms for customizing the retrieval of audiovisual content corresponding to an elapsed portion of an ongoing event based on the time of entry of a late-joining user thereby enabling adaptive recap summaries. Moreover, the content of the summaries can be further customized based on contextual information such as a user's role (e.g., a job title) thereby ensuring helpful recap summaries.
H04L 65/403 - Dispositions pour la communication multipartite, p. ex. pour les conférences
H04L 12/18 - Dispositions pour la fourniture de services particuliers aux abonnés pour la diffusion ou les conférences
H04L 51/043 - Messagerie en temps réel ou quasi en temps réel, p. ex. messagerie instantanée [IM] en utilisant ou en gérant les informations de présence
30.
PASSIVE PEN AND TOUCH PROFILING AND CLASSIFICATION FOR A CAPACITIVE TOUCH INTERFACE
A capacitive touch interface includes passive pen and touch profiling and classification. Distinguishing passive pen input from user touch improves the accuracy of input detection and expands access to pen-related features of the operating system and applications. Features extracted from a touch event dataset can include touch location, size, shape, duration, consistency, capacitance intensity, dynamics (e.g., shape change, size change, movement pattern), clustering, palm indication, pressure, etc. Datasets and features span multiple frames of a touch device. A classifier (e.g., a machine-learning (ML) model, such as a neural network) classifies each contact as finger, palm, passive pen, etc. based on the touch profile indicated by features. A post-processor can determine a final classification, e.g., by comparing a probability associated with the classification to a final classification threshold. Classification can take place over one or more frames to confirm classification before a handler responds to the classified touch events.
G06F 3/0354 - Dispositifs de pointage déplacés ou positionnés par l'utilisateurLeurs accessoires avec détection des mouvements relatifs en deux dimensions [2D] entre le dispositif de pointage ou une partie agissante dudit dispositif, et un plan ou une surface, p. ex. souris 2D, boules traçantes, crayons ou palets
G06F 3/041 - Numériseurs, p. ex. pour des écrans ou des pavés tactiles, caractérisés par les moyens de transduction
G06F 3/048 - Techniques d’interaction fondées sur les interfaces utilisateur graphiques [GUI]
G06F 3/044 - Numériseurs, p. ex. pour des écrans ou des pavés tactiles, caractérisés par les moyens de transduction par des moyens capacitifs
31.
INCREASING SPEED AND EFFICIENCY IN AI ACCELERATOR RACKS WITH A FIXED COMMUNICATION TOPOLOGY
The present disclosure generally relates to computer hardware, configurations, and methods for increasing the speed and efficiency in artificial intelligence (AI) accelerator racks with a fixed communication topology. For example, the computer hardware, configurations, and methods described herein replace the network switch typically included in an AI accelerator rack with crossbar switches (208, 210) that are added to trays (202a-202d) of AI accelerators within the rack. One or more embodiments can include a tray (202a-202d) of AI accelerators (206a-206d) with both a y-dimension crossbar switch (208) and an x-dimension crossbar switch (210). Both the y-dimension crossbar switch (208) and the x-dimension crossbar switch (210) may be configured to a fixed AI model communication topology that enable the AI accelerators (206a-206d) within the tray to communicate rapidly across the y-dimension crossbar switch (208) and the x-dimension crossbar switch (210) with very little power expenditure. Additionally, an AI accelerator rack can include additional z-dimension crossbar switches across trays of AI accelerators (206a-206d).
G06F 15/173 - Communication entre processeurs utilisant un réseau d'interconnexion, p. ex. matriciel, de réarrangement, pyramidal, en étoile ou ramifié
H04L 41/0893 - Affectation de groupes logiques aux éléments de réseau
Embodiments of the present disclosure include techniques for managing partitions of hardware resources in a computer system. In one embodiment, the present disclosure includes a plurality of processors and a plurality of peripheral devices coupled together over a switch fabric. A host for a partition may include a driver and a registry. An OS and the driver may communicate with the registry. State information stored in the driver and registry may be coordinated to recover from link failures and other path changes in the partition, for example.
G06F 11/14 - Détection ou correction d'erreur dans les données par redondance dans les opérations, p. ex. en utilisant différentes séquences d'opérations aboutissant au même résultat
33.
MEMORY ACCESS FAULT CLUSTERING WITHOUT MEMORY FOOTPRINT INCREASE
A device identifies a first page fault resulting from an access by a processor system to an invalid first page table entry (PTE). The device identifies a plurality of PTEs as a fault clustering unit, the plurality of PTEs being contiguous and including the first PTE. The device satisfies the fault clustering unit, including allocating a different memory page to each PTE in the plurality of PTEs and marking each PTE as valid. After a subsequent page fault, the device confirms the fault clustering unit, including determining whether a threshold number of PTEs in the plurality of PTEs were accessed after the first page fault. If the threshold is not met, the device trims the fault clustering unit, including deallocating a memory page from a PTE in the plurality of PTEs that was not accessed after the first page fault and marking the PTE as invalid.
Systems and methods for managing multi-tier key-value (KV) caching for a generative artificial intelligence (AI) model are described. A system includes a lower tier memory, a higher tier memory, and a KV cache manager to: (1) determine whether a key-value (KV) cache corresponding to a user associated with a session exists in the lower tier memory, and (2) if the KV cache, or a portion thereof, does not exist in the lower tier memory, upon determining that at least a portion of the KV cache exists in the higher tier memory, load at least the portion of the KV cache into the lower tier memory. The system includes a KV cache engine to execute: (1) an extended prefill phase to generate keys and values for extending the KV cache in the lower tier memory, and (2) a decode phase to generate output tokens until a stopping criteria is met.
G06F 12/0802 - Adressage d’un niveau de mémoire dans lequel l’accès aux données ou aux blocs de données désirés nécessite des moyens d’adressage associatif, p. ex. mémoires cache
35.
HARDWARE DEBUG WITH RETRIEVAL-AUGMENTED GENERATIVE AI
The technology described herein relates to an AI based hardware debug system providing a method including receiving a test code script for a hardware functional block, tokenizing the test code script to generate a test code vector, indexing the test code vector for storage in a cloud-based test code vector database, inputting the indexed test code vector into a language model (LM), and in response to inputting the test code into the LM, receiving from the LM a suggestion of one or more related code scripts.
Systems and methods can be implemented to conduct a visual search as a service in a variety of applications. In various embodiments, a system is configured to provide searching capabilities of content provided by a first entity in response to a search request by a second entity. An image provided by the second entity can be used by the system as a query image to search the content of the first entity. In an embodiment, the first entity can be a commercial entity providing such a system with image related content regarding its products and services such that any number of individual consumers can search for particular products and services of the commercial entity via their communication enabled devices. In addition, such systems can be arranged for other embodiments to provide customized searches of a single source by many individual devices. Additional systems and methods are disclosed.
An agent may include a crawler engine configured to create a graph. An agent may include a state detector in communication with the crawler engine and configured to provide state information to the crawler engine. An agent may include an emulator in communication with the crawler engine, wherein the emulator generates emulated inputs in response to an action selected by the crawler engine.
The description relates to enhancing privacy of visual content displayed for a user. One example can identify an electronic privacy (E-privacy) filter capability on a display of a device. The example can surface the E-privacy filter capability for control by an entity presenting content on the display.
G06F 21/84 - Protection des dispositifs de saisie, d’affichage de données ou d’interconnexion dispositifs d’affichage, p. ex. écrans ou moniteurs
G06F 21/62 - Protection de l’accès à des données via une plate-forme, p. ex. par clés ou règles de contrôle de l’accès
G09G 5/00 - Dispositions ou circuits de commande de l'affichage communs à l'affichage utilisant des tubes à rayons cathodiques et à l'affichage utilisant d'autres moyens de visualisation
A screen sharing privacy manager performs method in response to initiating a screen sharing session on a computer network, in which the screen sharing session is of a first desktop environment of an operating system. A second desktop environment is generated with the operating system. The second desktop environment inaccessible to the screen sharing session. A selected item in the first desktop environment is moved to the second desktop environment based on a user profile.
This document relates to using machine learning models to automate the process of generating alert rules for computing systems. The disclosed techniques can generate templates that match log entries from system logs. Then, the templates can be employed to generate alert rules that specify certain conditions when alerts should be triggered. In some implementations, the machine learning model is invoked responsive to a determination that a given log entry does not already match an existing template. Thus, the number of invocations of the machine learning model is proportional to the number of templates, rather than the number of log entries. Since there are often many more log entries than templates, this approach can yield significant computational savings compared to approaches that might process each log entry using the machine learning model.
A computing device includes a processing component that generates heat during operation, and a device housing component proximate to the processing component, such that the device housing component is heated by the processing component. A plurality of fluid channels is thermally coupled with the device housing component. A fluid pump is configured to circulate cooling fluid through the plurality of fluid channels in the device housing component, such that the cooling fluid facilitates heat distribution across the device housing component.
Innovations in encoder-side options for intra block copy (“BC”) prediction mode facilitate intra BC prediction that is more effective in terms of rate-distortion performance and/or computational efficiency of encoding. For example, some of the innovations relate to concurrently performing block vector (“BV”) estimation and making block splitting decisions for a block. Other innovations relate to selectively merging blocks into a larger block during BV estimation.
H04N 19/159 - Type de prédiction, p. ex. prédiction intra-trame, inter-trame ou de trame bidirectionnelle
H04N 19/11 - Sélection du mode de codage ou du mode de prédiction parmi plusieurs modes de codage prédictif spatial
H04N 19/172 - Procédés ou dispositions pour le codage, le décodage, la compression ou la décompression de signaux vidéo numériques utilisant le codage adaptatif caractérisés par l’unité de codage, c.-à-d. la partie structurelle ou sémantique du signal vidéo étant l’objet ou le sujet du codage adaptatif l’unité étant une zone de l'image, p. ex. un objet la zone étant une image, une trame ou un champ
H04N 19/176 - Procédés ou dispositions pour le codage, le décodage, la compression ou la décompression de signaux vidéo numériques utilisant le codage adaptatif caractérisés par l’unité de codage, c.-à-d. la partie structurelle ou sémantique du signal vidéo étant l’objet ou le sujet du codage adaptatif l’unité étant une zone de l'image, p. ex. un objet la zone étant un bloc, p. ex. un macrobloc
H04N 19/57 - Estimation de mouvement caractérisée par une fenêtre de recherche de dimension ou de forme variables
H04N 19/593 - Procédés ou dispositions pour le codage, le décodage, la compression ou la décompression de signaux vidéo numériques utilisant le codage prédictif mettant en œuvre des techniques de prédiction spatiale
H04N 19/96 - Codage au moyen d'une arborescence, p. ex. codage au moyen d'une arborescence quadratique
43.
HYBRID MACHINE LEARNING MODEL ENVIRONMENT WITH HOMOMORPHIC ENCRYPTION
The technology described herein is related to a hybrid neural network that divides operations of a neural network layer between a server and a client device. In an aspect, one or more liner operations of a neural network layer are performed on the client, while non-linear operations, such as an activation function, are performed on the server. In an aspect, the technology described herein maintains network security by encrypting portions of the client-side components. The encrypted portions may be learned values, which may also be described as learned parameters. In aspects, homomorphic encryption is used.
A heat-sensing touch interface that profiles the temperature of touch data to identify user input is disclosed herein. Heat profiling touch input improves touch accuracy and allows users to interact with touch devices more naturally. Palm touches can be classified as non-user input, even when appearing as fragmented touches resembling finger touches, when thermal data associated with the palm touchpoint(s) exceeds the temperature range profile for finger touches. Moisture can be classified as non-user input when thermal data associated with the moisture touchpoint(s) is below the temperature range profile for finger touches. The temperature range of user input can be dynamically adjusted. Energy is conserved by activating or sampling a heat sensor array based on detection of touch data. Energy is also conserved by more accurately classifying touch inputs, resulting in reporting and processing fewer non-user inputs. Feedback can be provided at touchpoints classified as user input.
A method for optical multiplication includes displaying a first encoded electromagnetic signal on a shared memory space, the first encoded electromagnetic signal being based on an input value and displaying a second encoded electromagnetic signal on the shared memory space, the second encoded electromagnetic signal being based on a learned parameter associated with the input value. The method includes detecting a cumulative electromagnetic signal on the shared memory space including a combination of at least some of the first encoded electromagnetic signal and at least some of the second encoded electromagnetic signal and determining a product of the input value and the learned parameter based on decoding the cumulative electromagnetic signal.
Systems and methods are provided for determining a root cause of an incident that occurred in a 5G/6G multi-access edge computing and core network system. In particular, the disclosed technology is directed to using a pre-trained generative model to determine the root cause for an incident as recorded in event data of a system log. The present disclosure generates a prompt for the generative model to determine the root cause for an incident as recorded in the system log. The prompt comprises a combination including event data from a system log, function hierarchy graph data, and function information as grounding information in a prefix of the prompt. The prompt further includes a question that requests determining a root cause of an incident as recorded in the event data. An answer as generated by the pre-trained generative model indicates of the root cause and a function call associated with the incident.
H04L 41/0631 - Gestion des fautes, des événements, des alarmes ou des notifications en utilisant l’analyse des causes profondesGestion des fautes, des événements, des alarmes ou des notifications en utilisant l’analyse de la corrélation entre les notifications, les alarmes ou les événements en fonction de critères de décision, p. ex. la hiérarchie ou l’analyse temporelle ou arborescente
Devices, systems, and methods for memory-efficient post-quantum (PQ) cryptography are provided. A method includes generating, by a hardware security module (HSM), a cryptographic seed, storing, by the HSM, the cryptographic seed, receiving, by the HSM and from a cryptographic processor, a request for the cryptographic seed, and responsive to the request, providing, by the HSM and to the cryptographic processor, the cryptographic seed.
H04L 9/30 - Clé publique, c.-à-d. l'algorithme de chiffrement étant impossible à inverser par ordinateur et les clés de chiffrement des utilisateurs n'exigeant pas le secret
H04L 9/32 - Dispositions pour les communications secrètes ou protégéesProtocoles réseaux de sécurité comprenant des moyens pour vérifier l'identité ou l'autorisation d'un utilisateur du système
48.
PASSIVE PEN AND TOUCH PROFILING AND CLASSIFICATION FOR A CAPACITIVE TOUCH INTERFACE
A capacitive touch interface includes passive pen and touch profiling and classification. Distinguishing passive pen input from user touch improves the accuracy of input detection and expands access to pen-related features of the operating system and applications. Features extracted from a touch event dataset can include touch location, size, shape, duration, consistency, capacitance intensity, dynamics (e.g., shape change, size change, movement pattern), clustering, palm indication, pressure, etc. Datasets and features span multiple frames of a touch device. A classifier (e.g., a machine-learning (ML) model, such as a neural network) classifies each contact as finger, palm, passive pen, etc. based on the touch profile indicated by features. A post-processor can determine a final classification, e.g., by comparing a probability associated with the classification to a final classification threshold. Classification can take place over one or more frames to confirm classification before a handler responds to the classified touch events.
G06F 3/04883 - Techniques d’interaction fondées sur les interfaces utilisateur graphiques [GUI] utilisant des caractéristiques spécifiques fournies par le périphérique d’entrée, p. ex. des fonctions commandées par la rotation d’une souris à deux capteurs, ou par la nature du périphérique d’entrée, p. ex. des gestes en fonction de la pression exercée enregistrée par une tablette numérique utilisant un écran tactile ou une tablette numérique, p. ex. entrée de commandes par des tracés gestuels pour l’entrée de données par calligraphie, p. ex. sous forme de gestes ou de texte
G06F 3/0354 - Dispositifs de pointage déplacés ou positionnés par l'utilisateurLeurs accessoires avec détection des mouvements relatifs en deux dimensions [2D] entre le dispositif de pointage ou une partie agissante dudit dispositif, et un plan ou une surface, p. ex. souris 2D, boules traçantes, crayons ou palets
G06F 3/041 - Numériseurs, p. ex. pour des écrans ou des pavés tactiles, caractérisés par les moyens de transduction
G06F 3/044 - Numériseurs, p. ex. pour des écrans ou des pavés tactiles, caractérisés par les moyens de transduction par des moyens capacitifs
Systems, methods, devices, and computer readable storage media described herein provide techniques for encoding graph indexes. In an aspect, a request to store data in a dataset is received. A vector is determined based on the request and the vector is encoded into a prefix specifying a location of the vector in vector space. A first index key is generated from at least the prefix. The first index key is associated with a value of the vector in an index. In a further aspect, the index is a graph vector index. In another aspect, a second request to obtain data is received. A second index key is generated based on the second request. The second index key is utilized to locate the first index key in the index. The value of the data is returned as a response to the second request.
A vector communication engine enables vector-based communications by multiple AI agents by receiving a first vector message from a first AI agent. The first vector message embodies a vector representation format of the first AI agent corresponding to an embedding model used by the first AI agent to generate the first vector message. The first vector message is transmitted as a second vector message to a second AI agent, wherein the second vector message embodies a vector representation format of the second AI agent corresponding to an embedding model used by the second AI agent to process the second vector message.
An integrated circuit and method are disclosed for performing a remote direct memory access (RDMA) network operation with forward error correction (FEC) using an integrated circuit. A message is loaded from memory for transmission over a connection associated with a connection identifier. The message is logically divided into a first predetermined number of data blocks. A second predetermined number of parity blocks are determined based on the data blocks. The data blocks and parity blocks are transmitted using different source port numbers from a set of source port numbers associated with the connection identifier.
Some embodiments detect keys or other sensitive data in a document, and modify the document or surrounding data structures in order to prevent disclosure of protected data. Some modifications transform sensitive data inline, and some modifications replace the sensitive data with a reference to a copy of the sensitive data in a secure store. In some scenarios, the secure store is created specifically for the detected data. The document is not necessarily complete at the time of modification; some documents are streamed. Some embodiments base the modification at least partially on one or more of: a risk assessment, metadata embedded in the sensitive data, sensitivity level, trust level, intended recipient, policy, or regulatory requirements, for example.
Methods, apparatuses, and products for automatically routing incident tickets, including: receiving an incident ticket describing an event in a computing system; calculating a plurality of routing signals for the incident ticket, wherein calculating the plurality of routing signals comprises: calculating, based on a plurality of user-defined rule sets applied to the incident ticket, one or more rule-based signals of the plurality of routing signals; calculating, based on one or more machine learning models sets applied to the incident ticket, one or more model-based signals of the plurality of routing signals; and performing a routing decision for the incident ticket based on the plurality of routing signals.
Example solutions for providing route aggregation as a service (RaaS) are disclosed. A RaaS server receives original routing tables from routers in a wide area network (WAN), aggregates the routes by identifying common prefixes, and returns more compact aggregated routing tables to the routers. The routers use the aggregated routing tables to efficiently route data traffic in the WAN. Multiple resiliency approaches are disclosed, in order to prevent the routers from flushing the aggregated routes in the event that the RaaS server experiences an outage and the border gateway protocol (BGP session between each router and the RaaS server is lost. Example resiliency measures includes the use of a long lived stale time (LLST), reaching a length of perhaps weeks, during which the routers continue to use the aggregated routing tables, and the use of mirroring the aggregated routing tables in other RaaS
Techniques and systems are disclosed to optimize aspects of data protection, including with the use of homomorphic encryption and other privacy-protecting measures. Optimizing data protection of a data stream may include: segmenting the data stream into data portions; assigning sensitivity scores to the data portions based on a sensitivity of contents provided therein; selecting security levels for data protection of the data portions (e.g., encryption) based on a respective sensitivity score assigned to a respective data portion, as the security levels correspond to varying strengths of the data protection; determining the security levels to apply to the data portions, as the data portions are processed with combined data processing operations on the data portions associated with a same security level; and applying the data protection to the data portions by applying the varying strengths of the data protection to the data portions in accordance with the determined security levels.
The present disclosure relates to systems and methods for translating a structured query language (SQL) with a user-defined function (UDF) into a translated SQL query using a machine learning model. The systems and methods automate a verification of the translated SQL query establishing an equivalence between the translated SQL query and the UDF.
An annotation tool receives, from a website or application, a graphical user interface (GUI) definition and a GUI image generated by the website or application. The tool identifies an interactive element visible in the GUI image and corresponding code in the GUI definition for the interactive element and infers, from the GUI image and the GUI definition, a description of the interactive element, the description including, in natural language, a purpose of the interactive element. The tool generates an annotation for the interactive element, the annotation comprising the description and updates the website or application based on the annotation. In example scenarios, the annotation is used by an artificial intelligence agent for performing actions on the GUI without referring to the GUI image.
Systems and methods for generating a design difference caption are provided. In particular, a computing device may receive a capture request to capture a content item, capture the content item with metadata associated with the content item, and generate one or more semantic embeddings of content data using one or more semantic models. The content data may include the content item and the metadata. The computing device may further store the one or more semantic embeddings in a content database, provide captured content history stored in the content database, the capture content history including a list of captured content items, receive a user query to search the captured content history; and provide search results based on the user query.
Recommendations of access control assignments are disclosed. Paired activity data is determined based on historical activity. The paired activity data including task and resource pairs that indicate a task performed by an individual identity, and a resource on which the task was performed. Over-privileging metrics are determined for candidate access control assignments based on the tasks and resource scopes granted by the candidate access control assignments and the paired activity data. Recommended access control assignments are determined as a subset of the candidate access control assignments with a lowest aggregate over-privileging metric whose combined permissions and resource scopes cover at least a predetermined percentage of the paired activity data. A responsive action is performed based on the recommended access control assignments.
A method of providing data communication between a first device and a second device includes, establishing a first communication link with a downstream device connected to the second device using a first mode via a USB-type interface, wherein in the first mode the USB-type interface utilizes a first set of USB communication lanes; establishing a second communication link with the first device via the USB-C port using an Alternate mode wherein the Alt-mode utilizes the first set of USB communication lanes; and, in accordance with establishing the second communication link, changing a mode of the first communication link so that the first communication link does not communicate via the first set of USB communication lanes.
Systems and methods for managing multi-tier key-value (KV) caching for a generative artificial intelligence (AI) model are described. A system includes a lower tier memory, a higher tier memory, and a KV cache manager to: (1) determine whether a key-value (KV) cache corresponding to a user associated with a session exists in the lower tier memory, and (2) if the KV cache, or a portion thereof, does not exist in the lower tier memory, upon determining that at least a portion of the KV cache exists in the higher tier memory, load at least the portion of the KV cache into the lower tier memory. The system includes a KV cache engine to execute: (1) an extended prefill phase to generate keys and values for extending the KV cache in the lower tier memory, and (2) a decode phase to generate output tokens until a stopping criteria is met.
G06F 12/0868 - Transfert de données entre une mémoire cache et d'autres sous-systèmes, p. ex. des dispositifs de stockage ou des systèmes hôtes
G06N 3/063 - Réalisation physique, c.-à-d. mise en œuvre matérielle de réseaux neuronaux, de neurones ou de parties de neurone utilisant des moyens électroniques
62.
GENERATIVE LANGUAGE MODEL SYSTEM WITH EMPATHY-TUNED EXPERT ADAPTER
A computing system (10) is provided, including processing circuitry (12) configured to, in a training phase (20), receive a plurality of example multi-turn conversations (26) between users and generative language models involving tasks (28), where each task (28) has associated levels and/or types of empathy (30). The processing circuitry (12) is further configured to generate synthetic multi-turn conversations (41) for each task cluster (34) based on empathy patterns (38), using a generative language model system (50), and train an expert adapter (44) based on the synthetic multi-turn conversations (41) to adapt an output of the generative language model system (50) to achieve a task-specific target level and/or type of empathy. In an inference phase (24), the processing circuitry (12) is further configured to receive an inference-time user query (46) related to an inference-time task at the generative language model system (50), which includes a generative language model (48) and the trained expert adapter (44), and generate a response (54) to the inference-time user query (46) using the generative language model (48) and expert adapter (44).
A vector communication engine enables vector-based communications by multiple AI agents by receiving a first vector message from a first AI agent. The first vector message embodies a vector representation format of the first AI agent corresponding to an embedding model used by the first AI agent to generate the first vector message. The first vector message is transmitted as a second vector message to a second AI agent, wherein the second vector message embodies a vector representation format of the second AI agent corresponding to an embedding model used by the second AI agent to process the second vector message.
Methods, computer systems, and computer-storage media are provided for facilitating distributed encoding in a remote desktop environment. In embodiments, a representation of graphical data is provided to a first video encoder associated with a video driver of a remote desktop server. The first video encoder may encode the graphical data to generate base-encoded graphical data. Thereafter, a representation of the base-encoded graphical data is provided to a second video encoder associated with a remote access protocol stack in a terminal service of the remote desktop server. The second video encoder may encode the base-encoded graphical data to generate fully-encoded graphical data. Such fully-encoded graphical data is transmitted to a remote desktop client device.
H04N 21/234 - Traitement de flux vidéo élémentaires, p. ex. raccordement de flux vidéo ou transformation de graphes de scènes du flux vidéo codé
H04N 21/60 - Structure du réseau ou procédés pour la distribution de vidéo entre le serveur et le client ou entre des clients distantsSignalisation de contrôle entre clients, serveur et éléments du réseauTransmission de données de gestion entre serveur et clientDétails de la communication entre serveur et client
A method of performing a collective operation using electromagnetic energy includes displaying a first encoded electromagnetic signal representing a first value encoded via a property of electromagnetic energy as a first pattern of one or more entries and displaying a second encoded electromagnetic signal representing a second value encoded via the property of electromagnetic energy as a second pattern of one or more entries. A first entry of the first pattern and a first entry of the second pattern are each displayed at a first shared entry window and are at least partially collocated. Based on detecting and counting the properties of electromagnetic energy at the first shared entry space, a count is determined for a first index position. The method includes determining a sum value of the first and second values based on the count and a multiplication factor for the first index position.
G06N 3/067 - Réalisation physique, c.-à-d. mise en œuvre matérielle de réseaux neuronaux, de neurones ou de parties de neurone utilisant des moyens optiques
G06N 3/084 - Rétropropagation, p. ex. suivant l’algorithme du gradient
66.
UTILIZING REGIONS OF INTEREST OF ELECTROMAGNETIC SIGNALS TO PERFORM COLLECTIVE OPERATIONS
A method of performing a collective operation using electromagnetic energy includes displaying a first encoded electromagnetic signal representing a first value encoded via a property of electromagnetic energy as a first pattern of one or more entries and displaying a second encoded electromagnetic signal representing a second value encoded via the property of electromagnetic energy as a second pattern of one or more entries. The method includes detecting, at a set of regions of interest, the property of electromagnetic energy of a first entry of the first pattern and of the second pattern and determining a count for a first index position based on counting the properties of electromagnetic energy detected for the first entry of the first pattern and of the second pattern. The method includes determining a sum value of the first and second values based on the count and a multiplication factor for the first index position.
G06N 3/067 - Réalisation physique, c.-à-d. mise en œuvre matérielle de réseaux neuronaux, de neurones ou de parties de neurone utilisant des moyens optiques
G06N 3/084 - Rétropropagation, p. ex. suivant l’algorithme du gradient
67.
USER INTERFACE CRAWLER AND AUTOMATED SOFTWARE TESTING SYSTEM
An agent may include a crawler engine configured to create a graph. An agent may include a state detector in communication with the crawler engine and configured to provide state information to the crawler engine. An agent may include an emulator in communication with the crawler engine, wherein the emulator generates emulated inputs in response to an action selected by the crawler engine.
Decoupling of telemetry data production from network transmission during network congestion in a processor-based system is disclosed. The processor-based system includes processor cores ("cores") that can each execute computer program instructions to perform tasks. To avoid the issue of analyzing telemetry data relating to the cores collected for a given epoch that may include missing and thus incomplete telemetry data, such as due to network congestion, a telemetry data system in the processor-based system is configured to track timing of receipt of telemetry data packets. If any telemetry data packets have not been timely received for an expected epoch, an epoch status for such epoch is recorded as having incomplete telemetry data for such epoch. This ensures complete telemetry data received for a given epoch is analyzed to more accurately control performance in the processor-based system.
Techniques are described herein that are capable of performing over-subscription ratio-based sharding in an AI system. The AI system includes a first rack of first AI accelerators, NICs associated with the first AI accelerators, and a second rack of second AI accelerators. Subsets of the first AI accelerators establish a consensus regarding allocation of a workload among the first AI accelerators by communicating via the NICs. The consensus indicates portions of data that are to be transmitted from the first rack to the second rack via a Tier1 switch and time durations to be consumed by transmission of the portions. The first AI accelerators group the portions into shards corresponding to respective timeslots based at least on the time durations and an over-subscription ratio associated with the first AI accelerators. Designated first AI accelerators transmit the shards in the respective timeslots to the second rack via the Tier1 switch.
The present disclosure relates generally to systems and methods for executing a search query using a combination of single-vector searching and multi-vector comparing to more efficiently and accurately identify a set of documents that are responsive to the search query. This involves applying a vector generation model to a search query to generate both single-vector and multi-vector embeddings for the search query. A single-vector search is performed to determine a candidate set of documents that might be relevant to the search query. A more robust multi-vector comparison is then performed to generate a result set including a re-ranked set of documents and/or a further reduced set of documents that is responsive to the search query.
09 - Appareils et instruments scientifiques et électriques
41 - Éducation, divertissements, activités sportives et culturelles
42 - Services scientifiques, technologiques et industriels, recherche et conception
Produits et services
downloadable software for a logic and puzzle game and social networking via a global computer network and wireless device; downloadable recreational logic and puzzle game software accessed via mobile phones, handheld computers and other wireless devices; downloadable mobile application for a logic and puzzle game for recreational gaming purposes; downloadable logic and puzzle game software via a global computer network and wireless devices providing entertainment information in the field of interactive logic and puzzle games and recreational logic and puzzle game playing; entertainment services, namely, providing a website featuring logic and puzzle games; entertainment services, namely, providing online electronic recreational logic and puzzle games; multimedia publishing of software, games, and electronic publications featuring interactive logic and puzzle games providing temporary use of online non-downloadable software for a logic and puzzle game for recreational purposes; providing temporary use of online non-downloadable software for logic and puzzle games and social networking; providing temporary use of online non-downloadable logic and puzzle game software
72.
DYNAMIC AI-ASSISTED GUIDANCE FOR ELECTRONIC RECORD COMPLETION
An electronic record input system (ERIS) may detect historical entities in the electronic record, the historical entities corresponding to the context and corresponding to first input fields of the electronic record. The ERIS may identify a next input field of the electronic record based on the historical entities by executing a predictive model on the historical entities, wherein the next input field corresponds to a next entity in the context, the next input field is different from the first input fields. The ERIS may generate one or more suggested actions for guiding the collection of the next entity corresponding to the next input field by executing a generative artificial intelligence model on the next input field and the historical entities. The ERIS may present the one or more suggested actions using the user computing device.
Tensile stress in a glass substrate of an optical storage medium is reduced by pre-stressing the substrate in compression to counteract thermally- and inertially-induced tensile stresses experienced during write operations using a pulsed laser. In a first embodiment, a uniform compressive stress distribution is realized in the medium by performing write operations in an enclosure that is positively pressurized by a gas. In a second embodiment using a disc-shaped optical storage medium, compressive forces are directed radially inward using circumferentially-located mechanical engagement members. A third embodiment uses a metal tension strip that is heat-treated to induce tensile hoop stress in the strip. The metal tension strip functions as a circumferential compression band around a disc-shaped optical storage medium. Axially-directed forces can be applied to support the center of the disc in the second and third embodiments to mitigate against buckling from the radially-inward directed forces.
G11B 33/14 - Diminution de l'influence des paramètres physiques, p. ex. changements de température, humidité, poussière
G11B 7/00 - Enregistrement ou reproduction par des moyens optiques, p. ex. enregistrement utilisant un faisceau thermique de rayonnement optique, reproduction utilisant un faisceau optique à puissance réduiteSupports d'enregistrement correspondants
G11B 7/0037 - Systèmes d'enregistrement, de reproduction ou d'effacement caractérisés par la forme du support avec des disques
Techniques are described herein that are capable of performing a multi-vector correlation using threat and anomaly analyses to assess a real-time security threat. A supervised AI model is caused to determine a first extent to which relationships between more than two vectors in a multi-vector representation, which indicate telemetry associated with a machine, correspond to first reference relationships associated with a known security threat. An unsupervised AI model is caused to determine a second extent to which the relationships differ from second reference relationships associated with a machine type of the machine. A security score is generated by combining indications of the first and second extents. A security action is performed as a result of the security score satisfying a security threat criterion.
H04L 41/16 - Dispositions pour la maintenance, l’administration ou la gestion des réseaux de commutation de données, p. ex. des réseaux de commutation de paquets en utilisant l'apprentissage automatique ou l'intelligence artificielle
Methods, apparatuses, and products for runtime firmware updates using hotpatching, including: storing, in a runtime buffer, a firmware update patch comprising an updated code segment for a firmware module loaded into system memory for execution; adding, to the firmware module and at a location corresponding to a code segment updated by the updated code segment, a redirect command directed to the updated code segment; and wherein the redirect command, when executed, causes the updated code segment to be executed instead of the code segment.
Systems and methods for resolving a performance bottleneck of a query search to improve search results are provided. In particular, a computing device may receive a user search query from a user, generate query alteration candidates based on the user search query using an index-based language model, determine whether the user search query includes a bottleneck term that is likely to cause the performance bottleneck of the query search, wherein the bottleneck term includes one or more words or phrases, optimize, in response to determination that the user search query includes the bottleneck term, the query alteration candidates to resolve the performance bottleneck, update the index-based language model by updating confidence scores associated with the query alteration candidates, generate search results based on the optimized query alteration candidates, and provide the search results to the user.
The described technology provides a method including selecting a physical SFT entry to be victimized, the physical SFT entry being one of a base SFT entry and an extra SFT entry of a logical SFT entry in a snoop filter (SFT) and CPU CPU GPU consolidating the base SFT entry and the extra SFT entry into one physical entry.
G06F 12/0831 - Protocoles de cohérence de mémoire cache à l’aide d’un schéma de bus, p. ex. avec moyen de contrôle ou de surveillance
G06F 12/0891 - Adressage d’un niveau de mémoire dans lequel l’accès aux données ou aux blocs de données désirés nécessite des moyens d’adressage associatif, p. ex. mémoires cache utilisant des moyens d’effacement, d’invalidation ou de réinitialisation
78.
SYSTEMS AND METHODS FOR TARGET COMMUNICATION ITEMS HAVING CONTEXTUAL TITLES DISTINCT FROM RELATED COMMUNICATION ITEMS
Technology is disclosed to programmatically determine and generate, for a target communication item, a contextual title that is relevant to a user, optimally formatted for ease of indexing and use by certain software applications, and easy for the user to consume. The contextual title is generated based on a comparison of the target communication item and related communication items. In this manner, the contextual title generated for the target communication item is different from existing contextual titles of the related communication items, yet related enough to associate the target communication item to certain related communication items. In this manner, contextual titles are generated without a user having to open or edit metadata associated with the communication item, thereby conserving bandwidth associated with a user engaging with high-bandwidth content from the target communication item and battery associated with consuming and editing the document.
Systems and methods are provided for implementing adaptable embedded search engine functionality. In an aspect, a shared SERP system receives a user search query from a first search utility among a plurality of search utilities, which is associated with corresponding apps that are different from each other. A router of the shared SERP system provides the user search query and location information to a first query builder among one or more query builders of the shared SERP system. The first query builder constructs a query request corresponding to the user search query, based on the provided user search query and location information. A first query executor among one or more query executors of the shared SERP system executes the query request to produce search results. A component renderer of the shared SERP system renders one or more UX components within the first SERP, based on the search results.
G06F 16/955 - Recherche dans le Web utilisant des identifiants d’information, p. ex. des localisateurs uniformisés de ressources [uniform resource locators - URL]
80.
TRACE RECONSTRUCTION FROM READS WITH INDETERMINANT ERRORS
Polynucleotide sequencing generates multiple reads of a polynucleotide molecule. Many or all of the reads contain errors. Trace reconstruction takes multiple reads generated by a polynucleotide sequencer and uses those multiple reads to reconstruct accurately the nucleotide sequence of the polynucleotide molecule. Some reads may contain errors that cannot be corrected. Thus, there may be reads that can be used throughout their entire length and other reads that have indeterminant errors which cannot be corrected. Rather than discarding the entire read when an indeterminant error is found, the portion of the read with the error is skipped and the sequence of the read following the error is used to reconstruct the trace. The amount of the read skipped is determined by the location of subsequence after the error that matches a consensus sequence of the other reads. Analysis resumes at a location determined by the location of the match.
G16B 40/00 - TIC spécialement adaptées aux biostatistiquesTIC spécialement adaptées à l’apprentissage automatique ou à l’exploration de données liées à la bio-informatique, p. ex. extraction de connaissances ou détection de motifs
This disclosure describes utilizing a generative document system to create generative search results documents using generative artificial intelligence (AI) models and dynamically determining which one of the generative search results documents to provide in response to a search query. For example, in response to receiving a search query, the generative document system obtains search link results (e.g., website links and corresponding grounding information) for the search query and utilizes this information with multiple generative AI models to generate various types of generative search results documents. Additionally, the generative document system generates and utilizes a generative document arbitration model to determine, based on the search link results, which of the generative search results documents to provide in response to the search query.
Searching heterogeneous cybersecurity data sources is described. A natural language search enables searching across available data sources, independent of underlying storage architectures or query languages. This capability leverages multiple task-specific agents, large language models, and data map representation-based approaches, and enables integration of new data sources into the search process with minimal additional computing resources. The searching includes determining a data map representation for available data sources (which might have different storage architectures or query languages), which indicates relationships between the available data sources. Query intent associated with a query is determined. A query intent task is determined by mapping the query intent to a relevant data source according to the data map representation. A search agent parameterized for searching the relevant data source executes the query intent task. The result of this executed query intent task is used to augment a response to the query.
Implementations for capturing user activity as memory for question-answering are provided. One example includes a computing system (100) comprising: processing circuitry (102) and memory (104) storing instructions that, during execution, cause the processing circuitry (102) to: receive a query (130); perform semantic searching on an L2 index (118) comprising a database of event summaries (400) based on the query (130) to retrieve an L2 answer (500) comprising one or more relevant event summaries; generate one or more L1 answers (502) based on the L2 answer (500) using an L1 index (110) comprising a database of screen summaries (300) associated with a plurality of L0 screenshot images (106), wherein each of the one or more L1 answers (502) comprises relevant screen summaries (300) corresponding to L0 screenshot images (106) associated with a respective relevant event summary; pass the query (130) and the one or more L1 answers (502) to a large language model (120) to generate an answer (132) to the query (130); and output the answer (132) to the query (130).
The description relates to enhancing component and device reliability by analyzing stored margin data relating to the component. One example can obtain first margin data relating to a component from a first time and obtain second margin data relating to the component from a subsequent second time. The example can compare the first margin data to the second margin data to identify a change in a condition of the component.
A computerized method detects and mitigates regression errors in software services. Error log data associated with a current version of the service is obtained from a data store. The error log data is transformed into a first group of error log vectors using a model, wherein the first group of error log vectors represents semantic meaning of the error logs. A second group of error log vectors associated with a previous version of the service is obtained. Based on semantic similarity, the first group and the second group of error log vectors are clustered and, based on the clustering, a vector cluster that includes error log vectors from the first group and no error log vectors from the second group is identified as a regression vector cluster. Regression data associated with the regression vector cluster is generated and a regression mitigation action is performed using the generated regression data.
Efficiently replacing software breakpoint instructions in processor-based devices is disclosed. In this regard, in one exemplary embodiment, a processor-based device is provided. The processor-based device comprises a system memory and a processor. The processor comprises a breakpoint slip register (BSR) configured to store an instruction of a software process that was replaced in the system memory by a software breakpoint instruction, and further comprises a breakpoint slip enable (BSE) indicator. The processor is configured to, during execution of the software process, execute the software breakpoint instruction. The processor is further configured to, responsive to executing the software breakpoint instruction, transfer program control to a debugger. The processor is also configured to, upon program control returning from the debugger, determine that the BSE indicator is set. The processor is additionally configured to, responsive to determining that the BSE indicator is set, execute the instruction stored in the BSR.
The description relates to enhancing component and device reliability by analyzing stored margin data relating to the component. One example can obtain first margin data relating to a component from a first time and obtain second margin data relating to the component from a subsequent second time. The example can compare the first margin data to the second margin data to identify a change in a condition of the component.
This invention relates to a method and system for Active Multi-Preference Optimization (AMPO) in training language models. The method involves generating multiple candidate responses for a given query and selecting a subset of responses that are most informative for training. The selection process includes identifying a highest-reward response as a positive example and employing strategies such as clustering and weighted selection to choose negative responses. The selected responses are used to compute a weighted contrastive loss function, which updates the model's parameters to improve alignment with human preferences. This approach enhances training efficiency by reducing computational overhead and improving model performance through targeted data selection and optimization.
The techniques describe dynamic content-aware data sampling that adjusts a sampling rate of data associated with a service based on the content of the data. More specifically, if the data to be logged indicates that the service is operating normally, then the sampling rate is adjusted lower as the logged data has limited purpose provided the service is likely not experiencing an issue. In contrast, if the data indicates that the service is operating abnormally, then the sampling rate is adjusted higher as the logged data has elevated purpose provided the service may be experiencing an issue. The dynamic content-aware data sampling is implemented using program code generated by an artificial intelligence agent (e.g., a generative language model) based on service owner requirements, as well as historical data patterns and/or usage patterns.
A data processing system implements prompting a first GenAI model to label a dataset into preference pairs based on a task and preference judgement criteria; using a second GenAI model to generate an initial reward model (RM) by modeling a reward score difference for each preference pair as a reward distribution curve that ranks each preference pair; training adapters of the second model to iteratively train the initial RM by: identifying on the curve a first portion including first pairs hard for the first model to correctly label, and a second portion including positive pairs ranked higher than the first pairs, and updating the curve for a subsequent iteration RM based on an improved dataset including the positive pairs and human annotated data generated from the first pairs; and fine-tuning the first model based on either a last iteration RM or an improved dataset associated with the last iteration RM.
A computerized method detects and mitigates regression errors in software services. Error log data associated with a current version of the service is obtained from a data store. The error log data is transformed into a first group of error log vectors using a model, wherein the first group of error log vectors represents semantic meaning of the error logs. A second group of error log vectors associated with a previous version of the service is obtained. Based on semantic similarity, the first group and the second group of error log vectors are clustered and, based on the clustering, a vector cluster that includes error log vectors from the first group and no error log vectors from the second group is identified as a regression vector cluster. Regression data associated with the regression vector cluster is generated and a regression mitigation action is performed using the generated regression data.
A method for calibrated touch sensing on an electronic device comprises: (a) scanning driven electrodes of a touch sensor while recording signal from sensory electrodes of the touch sensor; (b) storing the signal in a signal matrix; (c) for each element in the signal matrix that is lower than a predetermined threshold, adjusting a corresponding element in the calibration matrix based at least partly on a value of the element in the signal matrix and on a value of the corresponding element in the calibration matrix; (d) for each element in the signal matrix that is above the predetermined threshold, estimating the corresponding element in the calibration matrix based at least partly on one or more elements neighboring the corresponding element in the calibration matrix; and (e) returning a calibrated signal value based at least partly on corresponding values in the signal and calibration matrices.
Methods, computer systems, and computer-storage media are provided for facilitating distributed encoding in a remote desktop environment. In embodiments, a representation of graphical data is provided to a first video encoder associated with a video driver of a remote desktop server. The first video encoder may encode the graphical data to generate base-encoded graphical data. Thereafter, a representation of the base-encoded graphical data is provided to a second video encoder associated with a remote access protocol stack in a terminal service of the remote desktop server. The second video encoder may encode the base-encoded graphical data to generate fully-encoded graphical data. Such fully-encoded graphical data is transmitted to a remote desktop client device.
Decoupling of telemetry data production from network transmission during network congestion in a processor-based system is disclosed. The processor-based system includes processor cores (“cores”) that can each execute computer program instructions to perform tasks. To avoid the issue of analyzing telemetry data relating to the cores collected for a given epoch that may include missing and thus incomplete telemetry data, such as due to network congestion, a telemetry data system in the processor-based system is configured to track timing of receipt of telemetry data packets. If any telemetry data packets have not been timely received for an expected epoch, an epoch status for such epoch is recorded as having incomplete telemetry data for such epoch. This ensures complete telemetry data received for a given epoch is analyzed to more accurately control performance in the processor-based system.
H04L 47/26 - Commande de fluxCommande de la congestion utilisant un retour explicite à la source, p. ex. paquets de signalisation de congestion
H04L 43/022 - Capture des données de surveillance par échantillonnage
H04L 47/10 - Commande de fluxCommande de la congestion
H04L 47/129 - Prévention de la congestionRécupération de la congestion au point de destination final, p. ex. réservation des ressources du terminal ou de l’espace en mémoire tampon
95.
DETECTING AND PROCESSING INCONSISTENT SCHEDULING DATA
Some embodiments detect infeasible scheduling requests, such as overlapping do-not-move bookings with a given resource required at different locations, bookings with insufficient travel time, or other enumerated examples. Some embodiments utilize retrieval augmented generation or language model function invocation or both to summarize infeasibility patterns, or to explain scheduling decisions, or to do both. Some embodiments utilize virtual bookings to produce a partial schedule instead of declining to produce any schedule which involves conflicts. Some embodiments partition a scheduling problem into smaller problems using an eligibility graph, or using clusters computed from location data and shift assignment exploration, or both.
Systems and methods are provided for sorting cells with distinct cell designs for characterizing a library of the cell designs. The present disclosure uses randomized sorting rules associated with bins and pseudo-random numbers to counts the cells with measured fluorescence values in one of the bins. A mean fluorescence value for a cell design group may be determined based on a ratio of cell counts of the cells associated with the cell design group across the bins. Unlike the traditional histogram-based sorting that use a mean fluorescence value of a bin, the disclosed technology determines mean fluorescence values of cell design groups for characterizing libraries of the cell design group. Use of the mean fluorescence values with unbiased “sort-seq” and a de-multiplexed sequencing using the mean fluorescence values enables characterizing libraries of cell designs with improved accuracy over traditional use of discrete histograms.
Methods, systems, and computer storage media for providing data security posture management using an application discovery engine in a security management system. Application discovery supports identifying and mapping various applications within a computing environment. In particular, application discovery can be provided as part of security management operations to assess security posture of applications, identify vulnerabilities, and ensure compliance with regulations. In operation, application discovery data associated with a plurality computing resources of a computing environment is accessed. An annotated application discovery graph comprising a plurality of entities that represent the plurality of computing resources is generated. The annotated application discovery graph is deployed to support generating security postures for computing environments. A request is received for a security posture of the computing environment. A security posture visualization that includes an application discovery graph annotation is generated. The security posture visualization is communicated to cause display of the security posture visualization.
Examples of the present disclosure describe systems and methods for non-disruptively hibernating and resuming a guest environment using a network virtual service client. In examples, when a guest environment is hibernated, a network virtual service client provides an instruction to a virtual network interface card to set the device power state of the virtual network interface card to a low power state. The network virtual service client disables the communication channels used by the network virtual service client and saves the operating state of the virtual network interface card. When the guest environment is resumed, the network virtual service client provides an instruction to set the device power state of the virtual network interface card to a full power state. The network virtual service client reenables the communication channels used by the network virtual service client and restores the operating state of the virtual network interface card.
G06F 9/455 - ÉmulationInterprétationSimulation de logiciel, p. ex. virtualisation ou émulation des moteurs d’exécution d’applications ou de systèmes d’exploitation
99.
MEASURING PROBABILITY OF INFLUENCE USING MULTI-DIMENSIONAL STATISTICS ON DEEP LEARNING EMBEDDINGS
The disclosure herein describes a system for measuring probability of influence in digital communications to determine if communication content originated in a person's own prior knowledge or new information more recently obtained from interaction with communications of others. An estimated probability a new communication by a first user comes from the same distribution as prior communications of the first user are generated using multidimensional statistics on embeddings representing the communications. A second estimated probability that the new communication comes from the same distribution as communication(s) of a second user that were accessible to the first user are generated. If the second probability is greater than the first probability, the new communication is more likely influenced by exposure of the first user to the second user's communications rather than the first user's own historical knowledge. An influence attribution recommendation is generated, including an influence attribution or other recommended action.
G06Q 10/04 - Prévision ou optimisation spécialement adaptées à des fins administratives ou de gestion, p. ex. programmation linéaire ou "problème d’optimisation des stocks"
A server computing device generates training data based upon an identifier for a device, a timestamp, and a label received from a developer computing device. The server computing device trains a computer-implemented machine learning (ML) model based upon the training data. The server computing device also generates client configuration data for the ML model that specifies transformations that are to be applied to values in order to generate input values for the ML model. The server computing device deploys ML assets to client computing devices, the ML assets comprising the ML model and the client configuration data. The client computing devices execute the ML model using input values derived via transformations of (local) values produced by the client computing devices and transmit telemetry data to the server computing device. The server computing device updates the ML assets based upon the telemetry data.
