Abstract: Techniques and apparatus for mechanical assembly of an imaging device are described. An example image sensor assembly disposed within an imaging device includes a circuit board, an image sensor disposed on the circuit board, and an optics holder. The optics holder includes a support location, a thermally conductive material, and a plurality of extended surfaces that extend outward in a direction away from the circuit board. Each of the plurality of extended surfaces includes the thermally conductive material. The image sensor assembly also includes an optical lens disposed in the support location of the optics holder. The circuit board is disposed between the optics holder and a housing of the imaging device.

Abstract: This disclosure describes building a graph that is based on which IP address sources are active in the same time periods. By creating a node for each IP address with some moderate level of suspicious activity and connecting the nodes for any IP addresses that are engaging in that activity at roughly the same time, the graph represents scenarios of possible cooperation between sets of IP addresses. Nodes in this graph and the links between are assigned values based on how much suspicious activity they engage in during each time window. These values decay over time if no new suspicious activity occurs. By analyzing the resulting graph at any particular point in time, how similar two IP addresses have been behaving recently can be measured without having to keep track of their total history of activity over the last time period.

Abstract: An adapter layer may be used to customize a machine learning component by transforming data flowing into, out of, and/or within the machine learning component. The adapter layer may include a number of neural network components, or “adapters,” configured to perform a transformation on input data. Neural network components may be configured into adapter groups. A router component can, based on the input data, select one or more neural network components for transforming the input data. The input layer may combine the results of any such transformations to yield adapted data. Different adapter groups can include adapters of different complexity (e.g., involving different amounts of computation and/or latency). Thus, the amount of computation or latency added by an adapter layer can be reduced for simpler transformations of the input data.

Abstract: A user provides a question to be answered from detailed, dense or otherwise complex documents to a processing system that converts the question to a structured query language query and generates an embedding from the question, augmented by temporal data, synopses, themes, or other relevant information or data. The embedding is compared to embeddings generated from documents of a knowledge base to identify documents that are relevant to the question, and to rank such documents for their relevance. Highly ranking documents are combined with the query and provided to a language model that returns an answer to the question. A source for the answer is identified in at least one of the documents. The answer and the identified documents are presented to the user.

Abstract: Techniques for synchronizing device group data across different device control applications are described. A device group may include two or more smart home devices that may be controlled/supported by different device control applications. A user may set up device groups, including the same devices, in the different device control applications. A system may include a group synchronization service that synchronizes (e.g., merges) the multiple device groups across the different device control applications. The groups may be synchronized based on matching group names, matching devices included in the group, etc. After the device groups are synchronized, modifications made by the user to a device group in one device control application may be propagated to other device control applications.

Abstract: Techniques for volume durable log sequence number movement in a multi-volume database environment are described. Blocks of database updates are persisted in an atomic and durable manner, where the blocks may include updates to a first volume, a second volume, or both. Backlinks from one block of updates to the previous block of updates may be set, in addition to backlinks between updates to the first volume and backlinks between updates to the second volume. Upon durably persisting a block, the block backlinks can be followed to help verify that all contiguous blocks—of changes to one or both volumes—have all been persisted. Thereafter, a volume durable log sequence number (VDL) pointer value can be updated to point to a known “last” durably persisted log update.

Abstract: Technologies directed to using at least three antennas and Ultra-Wideband (UWB) protocols are described. One method includes sending a first signal to a second wireless device with data that cause it to send a second signal at a plurality of times. The method receives copies of the second signal at a first time via a first pair of antennas and copies of the second signal at a second time via a second pair of antennas. The method determines first and second angle-of-arrival (AoA) values from the copies of the second signals at the first and second times, respectively. The method determines location information of the second wireless device, the location information including an identifier corresponding to a first field of view (FOV) corresponding to the first pair or a second FOV corresponding to the second pair and at least one of the first AoA value or the second AoA value.

Abstract: A computing system receives encrypted data that can be decrypted by a first secret to obtain data, wherein the first secret is securely stored by the system, determines that the data encodes a second secret and executable code usable to perform cryptographic operations, and run the executable code to perform the cryptographic operations. The first secret may be a one-time pad.

Abstract: Techniques for compensating signal data from first order high pass filters for exponential decay to extract and underlying step function associated with the signal data. The techniques include adjusting a time scale of the signal data using a charge amplifier. The techniques also include determining a dynamic neutral point for the signal data using an exponential decay function. The techniques further apply an inverse transfer function to remove decay from the signal data and generate a step function associated with the signal data.

Abstract: Techniques for machine learning (ML) model adaptation via segment replacement and student-teacher training are described. A model optimizer determines that a ML model sought to be deployed to an edge device includes a source segment that is not supported by the edge device. The model optimizer identifies a replacement segment that is equivalent to the source segment, constructs an adapted ML model by swapping in the replacement segment for the source segment, and trains the adapted ML model, optionally from the source ML model using a student-teacher training procedure. The trained adapted ML model is then deployed to the edge device, where it can be successfully run with minimal (if any) degradation of performance compared to the original source ML model.

Abstract: Techniques for generating a summary of text-based documents are described. A system may be configured to generate a summary based on context data. The system may receive different types of context data corresponding to a user input. The context data may be converted to a linearized representation so that it can be processed by a decoder along with a source document for which the summary is being generated.

Abstract: In a system that controls access to resources via tokens, a system includes an application that generates ephemeral authorization tokens from partner tokens, to increase confidentiality and security, in embodiments. Responsive to a request, received by an application provider, for a protected resource, a federated ID/authorization provider is caused to receive a request for access/ID tokens that the ID/authorization provider provides (in any of various ways) to the application. The application validates and stores the tokens, nests the access/ID tokens within an ephemeral token document having a unique ID and shortened expiration, encrypts the nested ephemeral token using at least resource-specific encryption and causes the encrypted nested token to be sent to the protected resource provider that decrypts and validates the ephemeral token, causes the Access/ID token(s) within the ephemeral token to be validated, and provides the protected resource for valid tokens.

Abstract: Techniques for exchanging compressed gradient data within a distributed system are disclosed. A set of gradients are computed at a first worker node of the distributed system using a neural network model and a set of weights associated with the neural network model. Each of the set of gradients having a value less than a threshold is clipped, resulting in non-clipped data elements and clipped data elements. A mapping indicating which of the set of gradients correspond to non-clipped data elements and which of the set of gradients correspond to clipped data elements is generated. Compressed data is generated based on the non-clipped data elements. The mapping and the compressed data are transmitted from the first worker node to a second worker node of the distributed system.

Abstract: Techniques for managing logical camera devices are provided. A system for managing logical camera devices may instantiate logical camera devices having different responsibilities. A first logical camera device may receive and pass image data and metadata to a first application that may perform operations using both the image data and the metadata. A second logical camera device may receive and pass the metadata without the image data to a second application that may perform operations using the metadata but not the image data, thereby preserving power, memory, and/or processor resources that may have unnecessarily been used to provide the image data (e.g., via a buffer stream) to the second application, which may have deleted the image data upon receiving it. Another logical camera device may be instantiated to control other logical camera devices.

Abstract: Techniques and apparatus for adaptively controlling an end-effector of a robotic arm are provided. The end-effector includes at least one articulated finger having multiple facets arranged on a surface of the at least one articulated finger. The robotic arm is moved to engage an item using the at least one articulated finger. At least one of an amount of force or an amount of torque applied to the multiple facets on the surface of the at least one articulated finger is determined while the item is engaged using the at least one articulated finger. At least one of a position and orientation of the at least one articulated finger is adaptively controlled, based on at least one of the determined amount of force or the amount of torque.

Abstract: A method is disclosed to automatically detect overhangs from images with depth taken around an object during a scan of the object. An overhang detector can use an intersection of three filters based on these images and depth data associated with the images. The first filter looks for negative depth gradients along a 2D projection of a gravity vector, which is generally a vertical axis for images taken using a portrait orientation. The second filter selects the depth gradients that are oriented towards the projection of the gravity vector. The third filter is a salient object detection mask computed from the image. An intersection of the three filters can then be used to obtain overhangs. The method can be implemented in real time with a User Interface (UI) directing a user of a location of the overhang so that an image below the overhang can be taken.

Abstract: Data sets may be enhanced to create data stores. Request to create data stores may be received. As part of performing the request to create a data store, items stored in an extensible data format may be identified for machine learning enhancement. Machine learning models may be applied to generate additional data from data in the items. The additional data may be added to extend the items and store the extended items in a new data store.

Abstract: Various systems and methods of tracking database partition change log dependencies are described herein. A system is disclosed herein, according to some embodiments. A change log for a database partition of the plurality of database partitions is stored to a log store. An indication that a split of the database partition into two or more child database partitions has occurred is received. Responsive to the split of the database partition, the change log is split into two or more corresponding child change logs. A portion of the change log is deleted responsive to a determination that a respective complete snapshot is stored for each child database partition corresponding to a child change log of the change log and a determination that a retention window for the portion of the change log has passed.

Abstract: Described herein is a computer-implemented method for registering selections of graphical elements on electronic displays using time-of-flight sensors. First sensor data that is output by a first time-of-flight sensor can be received by a computing device. The first time-of-flight sensor can be positioned in a first orientation with respect to an electronic display. A position of an object on a surface of the electronic display can be determined based at least in part on the first sensor data. A user selection of a graphical element displayed on the electronic display can be registered at the determined position.

Abstract: Techniques for allocating resources and generating resource allocation instructions are described herein. A model can generate a historical error for resources of a facility based on a first set of predicted resource imbalances and historical data for resources imbalances of the facility for a first set of previous time periods. The model can receive a second predicted resource imbalance for the resources of the facility associated with a first future time period. The model can receive real-time signals for states of the resources of the facility associated with a present time period. The model can generate an adjusted resource prediction for the resources of the facility based at least in part on the historical error, the second predicted resource imbalance, and the real-time resource signals. The adjusted resource prediction can be associated with the first future time period.