Abstract: Operations of a workload are assigned to physical resources of a physical device array. The workload includes a graph of operations to be performed on a physical device array. The graph of operations is partitioned into subgraphs. Partitioning includes at least minimizing the quantity of subgraphs and maximizing resource utilization per subgraph. A logical mapping of the subgraph to logical processing engine (PE) units is generated using features of the subgraph and tiling factors of the logical PE units. The logical mapping is assigned to physical PE units of the physical device array at least by minimizing network traffic across the physical PE units. The operations of the subgraph are performed using the physical PE units to which the logical mapping is assigned. This process enhances the computational efficiency of the array when executing the workload.

Abstract: The invention relates to a method for communication between a third-party component (4) on a user device (2) and a service component (10) in the cloud (3), wherein: the service component (10) is signed by a data ID (13); the third-party component (4) provides component data (20); the component data (20) are marked by the data ID (13) in order to generate marked component data (21); the marked component data (21) are transferred to the cloud (3); and the marked component data (21) are assigned to the service component (10) having the data ID (13).

Abstract: It is known that a multiplicity of devices receive data from the cloud and make specific changes of state on the basis of these data. An example that can be mentioned is switching on a lamp or changing a color of a light source.

Abstract: A method for sparse matrix multiplication comprises receiving a first block having M elements in a first dimension, and parsing the first block of M elements into a first set of B sub-blocks including MB elements in the first dimension. A first sparsity mask having S % sparsity is applied to the first block of elements, such that each of the first set of B sub-blocks has S % sparsity. A second block is received having M elements in a second dimension, and is parsed into a second set of B sub-blocks that include MB elements in the second dimension. A second sparsity mask having S?% sparsity is applied to the second block of elements, such that S?% of the second set of B sub-blocks have 100% sparsity and (100?S?)% of the second set of B sub-blocks have 0% sparsity. The first and second blocks are then matrix multiplied.

Abstract: A method for operating a machine learning model is presented. The machine learning model includes a plurality of sequential transformer blocks. The method comprises receiving input data at a transformer block and processing the input data via a mixture of experts layer. At an auxiliary classifier, a measure of perplexity of the processed input data is determined. Based on the determined measure of perplexity, one or more experts in a downstream transformer block that will subsequently process the input data are indicated. Weight matrices are then fetched for the indicated one or more experts.

Abstract: A method is presented for training a neural network. For a weight matrix having integer dimensions M1 in a first dimension and an integer dimension M2 in a second dimension, a first balanced sparsity mask is generated that is an N1 of M1 mask in the first dimension. The first balanced sparsity mask is applied to the weight matrix during inference. A second balanced sparsity mask is generated for a transpose of the weight matrix. The second balanced sparsity mask is an N2 of M2 mask in the second dimension. The second balanced sparsity mask is applied to the transpose of the weight matrix during backpropagation.

Abstract: A computing system is configured to implement a deep neural network comprising an input layer for receiving inputs applied to the deep neural network, an output layer for outputting inferences based on the received inputs, and a plurality of hidden layers interposed between the input layer and the output layer. A plurality of nodes selectively operate on the inputs to generate and cause outputting of the inferences, wherein operation of the nodes is controlled based on parameters of the deep neural network. A sparsity controller is configured to selectively apply a plurality of different sparsity states to control parameter density of the deep neural network. A quantization controller is configured to selectively quantize the parameters of the deep neural network in a manner that is sparsity-dependent, such that quantization applied to each parameter is based on which of the plurality of different sparsity states applies to the parameter.

Abstract: A method is presented for operating a machine learning model including one or more mixture of experts layers. The method comprises receiving one or more input data shards at a routing gate network for a mixture of experts layer comprising a plurality of neural network experts. One or more neural network experts in the mixture of experts layer is designated layer to evaluate each input data shard. For each designated neural network expert, a weight matrix is retrieved having a predetermined sparsity to generate a sparsified designated neural network expert. Each input data shard is evaluated with a respective sparsified designated neural network expert.

Abstract: Embodiments described herein are directed to training techniques to reduce the power consumption and decrease the inference time of an NN. For example, during training, an estimate of power consumed by AMACs of a hardware accelerator on which the NN executes during inferencing is determined. The estimate is based at least on the non-zero midterms generated by the AMACs and the precision thereof. A loss function of the NN is modified such that it formulates the non-zero midterms and the precision thereof. The training forces the modified loss function to generate a sparse bit representation of the weights of the NN and to reduce the precision of the AMACs. Noise may also be injected at the output of nodes of the NN that emulates noise generated at an output of the AMACs. This enables the weights to account for the intrinsic noise that is experienced by the AMACs during inference.

Abstract: The invention relates to a method for certifying an application-specific cryptographic key in a certificate exchange service (30), comprising: receiving (130) a cryptographic attestation certificate (22) for an application-specific public key from an application (20) in an apparatus (10); checking (34; 136) the validity of the attestation certificate (22); and, if the attestation certificate (22) has been recognized as valid, comparing (34; 138) at least some information that has been extracted from the attestation certificate (22) with predefined reference information, and if the comparison reveals that a new certificate should be created, forming (36; 140) a new application-specific certificate (24) that comprises at least the application-specific public key extracted from the attestation certificate (22) and at least some of the information from the attestation certificate; transmitting (150) the new application-specific certificate (24) to the application (20), and to a method for requesting such certific

Abstract: The invention relates to a method for identifying an application (12) that is executed in an apparatus (10) to another communication participant, comprising: obtaining (200) a connection request for a secure connection between the application (12) and the other communication participant (30); forming (202) an information element (60) that comprises at least one item of information about the application (12); signing (204) the information element with a first secret key (52), which is part of a cryptographic asymmetric key pair that is certified by an information certificate (50) issued by an external trusted authority; incorporating the signed information element (60) into a connection request message (70), signing the connection request message with a secret device-specific key that is part of a cryptographic asymmetric key pair that is certified by a device-specific certificate of the apparatus, and transmitting the connection request message to the other communication participant.

Abstract: An evaluation device for evaluating an input signal (7), wherein the evaluation device has a base network (11), wherein the base network (11) is produced by a machine learning system and has an input layer (2) and a boundary layer (13), wherein the input layer (2) and the boundary layer (13) have a plurality of layers (4) arranged between them that are connected by means of connections (6), wherein the base network (11) is trained for a basic purpose, having at least two special networks (12, 12a-12b), wherein the special networks (12, 12a-12b) each form a machine learning system and each have a special network input layer (14a, 14b) and a special network output layer (15a, 15b), wherein the special networks (12, 12a-12b) are each trained and/or trainable for a special purpose.

Abstract: Systems and methods for extracting and sharing application-related user data are disclosed. A method may include extracting in-app data for at least one of the plurality of apps running on a computing device, the in-app data including content consumed by a user while the at least one app is running, and/or at least one user action taken in connection with the content. Using an entity template associated with the app, a plurality of text strings within the in-app data are classified into at least one of a plurality of data types specified by the template. At least one user data item (UDI) may be generated by combining at least a portion of the classified plurality of text strings, the at least one UDI being accessible by a second app, an operating system running on the, a service of the operating system, and/or a service running on at least another device.

Abstract: A system for providing online-content includes a data-acquisition module configured to collect user data and product data. A configuration module determines a predictive algorithm and one or more filters based on received marketer-input. A recommendation engine generates recommendations in real-time, in response to a request. The recommendations are generated via the predictive algorithm, which generates scalar scores from the dot-product multiplication of at least one user-vector and at least one product-vector, the scalar score indicating the likelihood that a desired interaction will occur between the user and the product. The algorithm also determines a hierarchical list based on the scalar scores, and applies the filters to the hierarchical list so as to identify the recommendations. The recommendation engine can generate recommendations in two ways: recommend items based on users, and recommend users based on items.

Abstract: Efficient encoding of high fanout communication patterns in computer programming is achieved through utilization of producer and move instructions in an instruction set architecture (ISA) that supports direct instruction communication where a producer encodes identities of consumers of results directly within an instruction. The producer instructions may fully encode the targeted consumers with an explicit target distance or utilize compressed target encoding in which a field in the instruction provides a bit vector for one-hot encoding. A variety of move instructions target different numbers of consumers and may also utilize full or compressed target encoding. In consumer paths where a producer is unable to target all consumers, a compiler may utilize various combination of producer and move instructions, using full and/or compressed target encoding to build a fanout tree that efficiently propagates the producer results to the all the targeted consumers.

Abstract: A system for generating recipient-targeted electronic messages includes an electronic message generation module and a user interface that allows a user to interact with the electronic message generation module to generate the electronic messages. The electronic messages may include a live content element, which provides content to the electronic message that is determined by detected circumstances of the electronic message being opened. The electronic messages may include a smart content element, which provides content to the electronic message that is determined by a record of community activity. The electronic messages may include a kinetic content element, which provides content to the electronic message that is interactive to the recipient of the electronic message.

Abstract: The automatic generation of one or more task-oriented conversational bots is disclosed. Illustratively, systems and methods are provided that allow for tracing the interactions of one or more computing applications inclusive of the interaction with one or more programmatic elements of the one or more computing applications, interaction with the graphical user interface(s) of the one or more computing applications, and/or the operation of the one or more computing environments on which the one or more computing applications are executing to collect various state data. The state data can be illustratively graphed to show the overall execution paths of one or more functions/operations of the one or more computing applications for use in generating one or more instructions representative of a desired task-oriented conversational bot that can be operatively executed through one or more application program interfaces of the one or more computing applications.

Abstract: Various technologies described herein pertain to prefetching content units. A prefetch request is transmitted to a server from a client device. The prefetch request includes data indicative of probabilities of slots for content units being available during an upcoming time period. The probabilities can be based on likely interaction with application(s) executed by the client device during the upcoming time period. Prefetched content units assigned to the client device for the upcoming time period can be received from the server responsive to the prefetch request. One or more of the prefetched content units can be served for display on a display screen of the client device during execution the application(s). Further, statuses of the prefetched content units can be monitored, and information that specifies a subset of the prefetched content units that are unlikely to be displayed on the display screen prior to corresponding deadlines for expiration can be transmitted.

Abstract: The present invention relates to a communications system (1) with at least one node (101, 102, 110), a control instance (10) and a shared communications medium (11). The at least one node is designed to receive messages that are or were transmitted in a first transmission mode. This first transmission mode has a lower data transmission rate and/or transmission complexity than a second transmission mode. The control instance (10) determines a first of the nodes in order to transmit data thereto in the second transmission mode or in order to release the shared communications medium (11) for the first node to transmit data in the second transmission mode. The control instance communicates this to the first node by means of a message (501, 502, 503, 504). This message is transmitted in the first transmission mode.

Abstract: Dynamic power routing is utilized to route power from other components, which are transitioned to lower power consuming states, in order to accommodate more efficient processing of computational tasks by hardware accelerators, thereby staying within electrical power thresholds that would otherwise not have accommodated simultaneous full-power operation of the other components and such hardware accelerators. Once a portion of a workflow is being processed by hardware accelerators, the workflow, or the hardware accelerators, can be self-throttling to stay within power thresholds, or they can be throttled by independent coordinators, including device-centric and system-wide coordinators.