Abstract: A method for detecting short-lived repeat voice commands includes activating a set of one or more warm words each associated with a respective action for controlling a long-standing operation performed by a digital assistant. While the digital assistant is performing the long-standing operation, the method includes receiving audio data, and detecting, in the audio data, a warm word from the activated set of one or more warm words. In response to detecting the warm word, the method includes performing the respective action associated with the detected warm word, and activating a set of one or more repeat warm words associated with the detected warm word. The method further includes receiving additional audio data, detecting, in the additional audio data, a repeat warm word from the activated set of one or more repeat warm words, and performing the respective action associated with the detected repeat warm word.

Abstract: A method using multi-assistant warm words includes, for each digital assistant in a group of digital assistants enabled on a multi-assistant device, receiving a respective active set of warm words that each specify a respective action to perform. Based on the respective active set of warm words, the method also includes executing a warm word arbitration routine to enable a final set of warm words for detection, each warm word in the final set of warm words selected from the respective active set of warm words for at least one digital assistant. While the final set of warm words are enabled for detection, the method includes receiving audio data corresponding to an utterance, detecting a warm word from the final set of warm words, and instructing the digital assistant associated with the detected warm word to perform the respective action specified by the detected warm word.

Abstract: This disclosure describes apparatuses and techniques for a switched multi-cell battery system for electronic devices. In some aspects, a switched multi-cell battery system may transfer, via a plurality of power control switches electrical power from a power adapter to components of the electronic device by charging battery cells in series and by discharging the battery cells in parallel or as a single battery cell. As a result, the switched multi-cell battery system may reduce or eliminate a voltage step-down conversion stage to increase a power-transfer efficiency of an electronic device. By doing so, charging times may be reduced or operating times may be increased, thereby improving users' experience with their electronic devices.

Abstract: Replicas are selected in a large distributed network, and the roles for these replicas are identified. In one example, a leader is selected from among candidate computing clusters. To make this selection, an activity monitor predicts or monitors the workload of one or more clients. Different activities of the workload are given corresponding weights. The delay in performing requested activities, modified by these weights is found, and the candidate leader with the lowest weighted delay is selected as the leader.

Abstract: A method includes obtaining a source training dataset that includes a plurality of source training images and obtaining a target training dataset that includes a plurality of target training images. For each source training image, the method includes translating, using the forward generator neural network G, the source training image to a respective translated target image according to current values of forward generator parameters. For each target training image, the method includes translating, using a backward generator neural network F, the target training image to a respective translated source image according to current values of backward generator parameters. The method also includes training the forward generator neural network G jointly with the backward generator neural network F by adjusting the current values of the forward generator parameters and the backward generator parameters to optimize an objective function.

Abstract: This application is directed to a display assistant device that acts as a voice-activated user interface device. The display assistant device includes a base, a screen and a rear speaker. The base is configured for sitting on a surface. The screen has a rear surface and is supported by the base at the rear surface. A bottom edge of the screen is configured to be held above the surface by a predefined height, and the base is substantially hidden behind the screen from a front view of the display assistant device. The rear speaker is concealed inside the base and oriented to project sound towards the rear side of the base.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer-readable media, for speech recognition using multi-dialect and multilingual models. In some implementations, audio data indicating audio characteristics of an utterance is received. Input features determined based on the audio data are provided to a speech recognition model that has been trained to output score indicating the likelihood of linguistic units for each of multiple different language or dialects. The speech recognition model can be one that has been trained using cluster adaptive training. Output that the speech recognition model generated in response to receiving the input features determined based on the audio data is received. A transcription of the utterance generated based on the output of the speech recognition model is provided.

Abstract: A method for training a non-autoregressive TTS model includes receiving training data that includes a reference audio signal and a corresponding input text sequence. The method also includes encoding the reference audio signal into a variational embedding that disentangles the style/prosody information from the reference audio signal and encoding the input text sequence into an encoded text sequence. The method also includes predicting a phoneme duration for each phoneme in the input text sequence and determining a phoneme duration loss based on the predicted phoneme durations and a reference phoneme duration. The method also includes generating one or more predicted mel-frequency spectrogram sequences for the input text sequence and determining a final spectrogram loss based on the predicted mel-frequency spectrogram sequences and a reference mel-frequency spectrogram sequence. The method also includes training the TTS model based on the final spectrogram loss and the corresponding phoneme duration loss.

Abstract: Techniques for controlling a remotely controllable device are described. In an example, a mobile device detects a remotely controllable device, measures a distance and direction from the mobile device to the remotely controllable device, and determines from the distance and direction that the mobile device is pointing at the remotely controllable device. In response to determining that the mobile device is in a handheld position, is pointing at the remotely controllable device, or both, the mobile device monitors for a movement of the mobile device according to a prescribed gesture. In response to detecting that the mobile device was moved according to the prescribed gesture, the mobile device presents a collection of selectable actions control operations of the remotely controllable device.

Abstract: This document describes techniques and systems for authentication management through IMU and radar. The techniques and systems use inertial sensor data from an inertial measurement unit (IMU) and/or radar data to manage authentication for a computing device. By so doing, the techniques conserve power, improve accuracy, or reduce latency relative to many common techniques and systems for computing-device authentication.

Abstract: Methods, systems, apparatus, including computer programs encoded on computer storage media, for reclaiming storage space in a storage environment. In one aspect, the method includes actions of aggregating data that is indicative of access to one or more data objects, determining a future storage cost associated with each of a plurality of data objects, determining an access window for each of the plurality of data objects, identifying a data object based on (i) the future storage cost that satisfies a predetermined threshold and (ii) a data object access window, providing a notification to a user device that requests feedback from a user indicating whether the data object can be deleted, and in response to receiving data that indicates that the data object can be deleted, generating an instruction to cause deletion of the data object upon the expiration of the access window.

Abstract: This document describes systems and techniques for facilitating spatial rediscovery using on-device hardware. For example, an audio input mechanism associated with a compact device can be activated and an audio signal can be generated by an audio output mechanism associated with an external device. The audio signal can be received at the audio input mechanism associated with the compact device. The receipt of the audio signal may cause the audio input mechanism to produce electrical signals having encoded information associated with the audio signal. The electrical signals can then be transmitted to the external device as wireless signals. Based on an analysis of the transmitted wireless signals, a spatial positioning of the compact device relative to the external device can be determined. Based on the determination of the spatial positioning, a location indicator can be provided via the external device.

Abstract: A computer-implemented input-method editor process includes receiving a request from a user for an application-independent input method editor having written and spoken input capabilities, identifying that the user is about to provide spoken input to the application-independent input method editor, and receiving a spoken input from the user. The spoken input corresponds to input to an application and is converted to text that represents the spoken input. The text is provided as input to the application.

Abstract: A method for augmenting handling of logs generated in platform as a system (PaaS) environments includes transmitting, to an external cloud computing environment, an application programming interface (API) request. The API request includes a trace identification (ID), a concealment indicator, and a policy ID. The method also includes updating the first entry in the first log corresponding to the API request based on the trace ID and the concealment indicator. The method also includes storing the first entry in the first log corresponding to the API request based on the storage criteria of the policy ID.

Abstract: A first electronic device, electronically coupled to a second device for supplying a charge to the second electronic device, tracks the voltage requirements of the second device and dynamically adjusts its output voltage upwards or downwards to match such requirements. The second electronic device may provide feedback to the first electronic device through a feedback loop. The feedback may include an indication of the voltage requirements and/or instructions for adjusting the voltage output of the first electronic device. The second device may be, for example, a wearable audio device, while the first device is a case for the wearable audio device.

Abstract: Methods and apparatus related to ranking modifications of a previous query. For example, modifications of a previous query may be generated based on a current query issued subsequent to the previous query by substituting one or more n-grams of the previous query with one or more n-grams of the current query. One or more measures of each of the modifications may be identified and, based on such measures, a ranking of each of the modifications may be determined. One of the modifications may be selected as a submission query based on the rankings of the modifications. The submission query may be selected for submission in lieu of, or in addition to, the current query.

Abstract: In some embodiments, a method for controlling access to media is provided, the method comprising: receiving an encrypted media content stream from a media content source that includes first media content corresponding to at least a portion of a media content item; requesting a second media content stream that includes second media content having a playback position adjacent to the first media content, wherein the second media content stream includes key information for decrypting the encrypted media content stream; decrypting the encrypted media content stream using the key information included in the second media content stream; and causing the decrypted media content stream to be presented.

Abstract: The present disclosure provides systems and methods for on-device machine learning. In particular, the present disclosure is directed to an on-device machine learning platform and associated techniques that enable on-device prediction, training, example collection, and/or other machine learning tasks or functionality. The on-device machine learning platform can include a mode controller that allows a user to provide data input indicating whether to operate one or more applications on the device in a first collection mode (e.g., permission mode) for storing training examples or a second collection mode for (e.g., incognito mode) for not storing training examples. The training examples can be generated based on user interaction with the one or more applications and used to personalize one or more machine-learned models used by the application(s) by retraining the models using the user-specific training examples.

Abstract: Implementations are provided for providing responsive audio recordings to user queries that are prerecorded by human beings, rather than generated automatically using speech synthesis processing. In various implementations, a query provided by a user at an input component of a computing device may be used to search a corpus of voice recordings From the searching, a plurality of candidate responsive voice recordings may be identified and ranked based on measures of credibility associated with speakers that created the candidate responsive voice recordings. Based on the ranking, one or more of the plurality of candidate responsive voice recordings may be provided for presentation to the user at an output component of the same computing device or a different computing device.

Abstract: Example embodiments of the present disclosure provide for an example method. The example method includes generating an initial user interface including a content assistant component. The example method include obtaining user input data. The example method includes processing, by a machine learned model interfacing with the content assistant component, the data indicative of the input received from the user. The method includes obtaining output data, from the machine learned model interfacing with the content assistant component, indicative of one or more content item components. The method includes transmitting data which causes the content item components to be provided for display via an updated user interface. The method includes obtaining data indicative of user selection of approval of the content item components. The method includes generating, in response to obtaining the data indicative of the user selection of the approval of the content item components, content items.

Abstract: Methods, systems, and apparatus include computer programs encoded on a computer-readable storage medium for delivering content. A method includes: identifying an application that is installed on a user device; accepting a bid from a sponsor associated with the application for presenting customized content on the user device after installation of the application on the user device; identifying an opportunity to present content on the user device including receiving a request for content that identifies the user device; conducting an auction including evaluating the accepted bid along with other bids; determining that the accepted bid is the winning bid in the auction and providing, in response to the determining, a shell for the customized content; enabling the shell to be populated with data stored on the user device to produce the customized content; and providing the customized content for presentation on the user device in response to the identified opportunity.

Abstract: A method is provided for processing medical text and associated medical images. A natural language processor configured as a deep conventional neural network is trained on a first corpus of curated free-text, medical reports each of which having one or more structured labels assigned by an medical expert. The network is trained to learn to read additional free-text medical reports and produce predicted structured labels. The natural language processor is applied to a second corpus of free-text medical reports that are associated with medical images. The natural language processor generates structured labels for the associated medical images. A computer vision model is trained using the medical images and the structured labels generated. The computer vision model can thereafter assign a structured label to a further input medical image. In one example, the medical images are chest X-rays.

Abstract: A method for determining playback behavior of a media item including a video component and an audio component includes determining a current event that causes a playing of the video component of the media item on a screen of a user device to be suspended. The method further includes predicting user playback preferences for the media item using a behavioral model trained on previous events corresponding to the current event, and determining, based on the predicted user playback preferences, whether to continue playing the audio component of the media item.

Abstract: Methods and systems of determining competition for search queries on a video hosting system on a computer network are provided. Search records associated with search queries directed to the video hosting system are received. Each search record identifies videos available on the video hosting system that are search results for a corresponding search query and indicates associated view durations that the videos were viewed responsive to the search query. Search queries directed to a first and second videos and associated view durations that the first video was watched in response to the queries are identified from the received search records. A competition score indicating an amount of competition between the first and second videos for views of users of the video hosting system is calculated based on the identified search queries and associated view durations.

Abstract: A method for example-driven machine learning is disclosed herein. The method comprises maintaining a plurality of dialog system rules and a knowledge database including a plurality of intent objects and a plurality of entity objects. The plurality of intent objects and the plurality of entity objects are associated with at least one dialog system rule. An exemplary phrase is received and one or more linguistic elements are retrieved from the exemplary phrase. It is determined that at least one of the linguistic elements is directed to at least one of the plurality of intent objects of the plurality of entity objects and the at least one of the linguistic elements in association with the at least one dialog system rule is added to the knowledge database.

Abstract: A system and method for repartitioning data in a distributed network. The method may include executing, by one or more processors, a first pass of a data set from a plurality of first sources to a plurality of first sinks, each first sink collecting data from one or more of the first sources, and executing, by the one or more processors, a second pass of the data set from a plurality of second sources to a plurality of second sinks, each one of the plurality of first sinks corresponding to one of the plurality of second sources, and each second sink collecting data from one or more of the second sources. Executing the first and second passes causes the data set to be repartitioned such that one or more second sinks collect data that originated from two or more of the first sources.

Abstract: A new segment of electronic handwriting is provided to a handwriting recognition module to obtain a plurality of textual interpretations of the new segment. The textual interpretations obtained from the handwriting recognition module are scored based on how each respective electronic handwriting representation would change a display of existing electronic content when the respective electronic handwriting representation is displayed substantially at the user designated position within or adjacent to the existing electronic content. Based on the scoring, an electronic handwriting representation corresponding to a respective textual interpretation of the plurality of textual interpretations is selected, and the existing electronic content is modified to include the selected electronic handwriting representation located substantially at the user designated position.

Abstract: Location modeling may be generated for candidate locations based on validated location data. The location models for candidate locations may be based on validated location data associated with wireless access points that is validated using transaction data. Such validation may include comparing an identifier associated with candidate location data received from a client device to an identifier for a transaction account associated with a transaction from transaction data at a candidate location. Matching transaction data and candidate location data may generate validated location data to be used in generating location models for candidate locations.

Abstract: Methods, systems, and media for associating multiple users with a media presentation device are provided. In some embodiments, the method comprises: receiving first user preferences; receiving second user preferences; receiving requests to associate the first user preferences and the second user preferences, respectively, with a media presentation device; causing the first user preferences and the second user preferences to be associated with identifying information of the media presentation device; receiving, from the media presentation device, a request to present a slideshow of images based on user preferences and identifying information of the first user device and a third user device; determining that the first user preferences are associated with the media presentation device; determining that user preferences of the third user device are not associated with the media presentation device; and causing content based on the first user preferences to be presented by the media presentation device.

Abstract: Techniques are described for selectively adapting and/or selectively utilizing a noise reduction technique in detection of one or more features of a stream of audio data frames. For example, various techniques are directed to selectively adapting and/or utilizing a noise reduction technique in detection of an invocation phrase in a stream of audio data frames, detection of voice characteristics in a stream of audio data frames (e.g., for speaker identification), etc. Utilization of described techniques can result in more robust and/or more accurate detections of features of a stream of audio data frames in various situations, such as in environments with strong background noise. In various implementations, described techniques are implemented in combination with an automated assistant, and feature(s) detected utilizing techniques described herein are utilized to adapt the functionality of the automated assistant.

Abstract: Implementations relate to automatic generation of speaker features for each of one or more particular text-dependent speaker verifications (TD-SVs) for a user. Implementations can generate speaker features for a particular TD-SV using instances of audio data that each capture a corresponding spoken utterance of the user during normal non-enrollment interactions with an automated assistant via one or more respective assistant devices. For example, a portion of an instance of audio data can be used in response to: (a) determining that recognized term(s) for the spoken utterance captured by that the portion correspond to the particular TD-SV; and (b) determining that an authentication measure, for the user and for the spoken utterance, satisfies a threshold. Implementations additionally or alternatively relate to utilization of speaker features, for each of one or more particular TD-SVs for a user, in determining whether to authenticate a spoken utterance for the user.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for automatically determining parameter values that control or affect provision of content by a content platform. In one aspect, evaluation points are identified for a parameter. Each evaluation point includes an evaluated parameter value of the parameter and a metric value of a metric corresponding to the provision of digital components by the content platform. A first model is generated using the set of evaluation points. A second model is generated based on the first model and an acquisition function that is based on mean values and confidence intervals of the first model and a configurable exploration weight that controls a priority of exploration for evaluating the parameter. A next parameter value to evaluate is determined from the second model and the content platform is configured to use the next parameter value to provide digital components.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for processing images using self-attention based neural networks. One of the methods includes obtaining one or more images comprising a plurality of pixels; determining, for each image of the one or more images, a plurality of image patches of the image, wherein each image patch comprises a different subset of the pixels of the image; processing, for each image of the one or more images, the corresponding plurality of image patches to generate an input sequence comprising a respective input element at each of a plurality of input positions, wherein a plurality of the input elements correspond to respective different image patches; and processing the input sequences using a neural network to generate a network output that characterizes the one or more images, wherein the neural network comprises one or more self-attention neural network layers.

Abstract: An electronic device is configured with a notification throttling system. The notification throttling system obtains a set of priorities for multiple applications that are installed on the electronic device and are enabled for alerting the user via notifications. The set of priorities specifies a relative importance of each particular application among the multiple applications. The notification throttling system throttles application alerts presented by the user device for notifications generated by the applications based on the relative importance of each particular application as specified by the set of priorities.

Abstract: The present disclosure provides systems and methods that generate query templates that are expressed in a generic schema-agnostic language. The query templates can be generated “from scratch” or can be automatically generated from existing queries, a process which may be referred to as “templatizing” the existing queries. As one example, generation of query templates can be performed through an iterative process that iteratively generates candidate templates over time to optimize a coverage over a set of existing queries. After generation of the schema-agnostic query templates, the systems and methods described herein can automatically translate/map the templatized queries into “concrete,” schema-specific queries that can be evaluated over specific customer schemas/datasets. In this manner, a query template for a given semantic query (e.g., “return the names of all employees”), is required to be written only once.

Abstract: Methods, systems, and computer programs encoded on a computer storage medium, for training and using machine learning models are disclosed. Methods include creating a model that represents relationships between user attributes, content exposures, and performance levels for a target action using organic exposure data specifying one or more organic exposures experienced by a particular user over a specified time prior to performance of a target action by the particular user and third party exposure data specifying third party exposures of a specified type of digital component to the particular user over the specified time period. Using the model, an incremental performance level attributable to each of the third party exposures at an action time when the target action was performed by the particular user is determined. Transmission criteria for at least some digital components to which the particular user was exposed are modified based on the incremental performance.

Abstract: This technology relates to rasterizing and compositing vector graphics in parallel on a data-parallel computing device. For example, vector data of the vector graphics, may be loaded into local memory accessible by the one or more parallel processors. The vector data may include one or more paths comprised of one or more path segments of the vector graphics. The one or more parallel processors may rasterize the one or more path segments into respective rasters and assign each of the rasters into groups based on pixel coordinates of the respective rasters. Each group may have an associated key and the rasters within each group represent a portion of the same vector graphic. The rasters may be placed onto subpixels according to their respective pixel coordinates and rendered onto a display by the one or more parallel processors.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for providing a machine learning model that is trained to perform a machine learning task. In one aspect, a method comprises receiving a request to train a machine learning model on a set of training examples; determining a set of one or more meta-data values characterizing the set of training examples; using a mapping function to map the set of meta-data values characterizing the set of training examples to data identifying a particular machine learning model architecture; selecting, using the particular machine learning model architecture, a final machine learning model architecture for performing the machine learning task; and training a machine learning model having the final machine learning model architecture on the set of training examples.

Abstract: This document describes techniques and apparatuses that enable determining expected hash-values in functions with control flow. A computing device receives a function comprising function instructions within at least three basic blocks connected via multiple execution paths. Hash-input instructions are inserted within a plurality of the basic blocks that indirectly force hash values at the respective insertion points. Hash values at ends of the plurality of the basic blocks are set to a canonical value and an expected hash-value and hash input-values are calculated using a hash function. By using the canonical value and the hash input-values, the expected hash-value is the same regardless of which execution path is executed.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for hotword detection on multiple devices are disclosed. In one aspect, a method includes the actions of receiving audio data that corresponds to an utterance. The actions further include determining that the utterance likely includes a particular, predefined hotword. The actions further include transmitting (i) data indicating that the computing device likely received the particular, predefined hotword, (ii) data identifying the computing device, and (iii) data identifying a group of nearby computing devices that includes the computing device. The actions further include receiving an instruction to commence speech recognition processing on the audio data. The actions further include in response to receiving the instruction to commence speech recognition processing on the audio data, processing at least a portion of the audio data using an automated speech recognizer on the computing device.

Abstract: A method includes receiving context biasing data that includes a set of unspoken textual utterances corresponding to a particular context. The method also includes obtaining a list of carrier phrases associated with the particular context. For each respective unspoken textual utterance, the method includes generating a corresponding training data pair that includes the respective unspoken textual utterance and a carrier phrase. For each respective training data pair, the method includes tokenizing the respective training data pair into a sequence of sub-word units, generating a first higher order textual feature representation for a corresponding sub-word unit, receiving the first higher order textual feature representation, and generating a first probability distribution over possible text units. The method also includes training a speech recognition model based on the first probability distribution over possible text units.

Abstract: The present disclosure describes thermal mitigation for an electronic speaker device and associated systems and methods. The thermal mitigation includes monitoring several thermal zones to determine or estimate thermal conditions in corresponding parts of the electronic speaker device. The thermal zones may include a System-on-Chip (SoC) integrated circuit (IC) component, audio components including power-dissipating IC components, and a temperature of an exterior surface of a housing component of the electronic speaker device. To mitigate thermal runaway, different throttling schemes may be triggered based on the thermal zones exceeding certain thermal limits. The throttling schemes may include reducing the amount of power supplied to the SoC, reducing audio power of the audio components to a lower wattage, or manipulating SoC cores such as by disabling one or more of the cores or adjusting utilization of the SoC cores.

Abstract: A method for extracting entities comprises obtaining a document that includes a series of textual fields that includes a plurality of entities. Each entity represents information associated with a predefined category. The method includes generating, using the document, a series of tokens representing the series of textual fields. The method includes generating an entity prompt that includes the series of tokens and one of the plurality of entities and generating a schema prompt that includes a schema associated with the document. The method includes generating a model query that includes the entity prompt and the schema prompt and determining, using an entity extraction model and the model query, a location of the one of the plurality of entities among the series of tokens. The method includes extracting, from the document, the one of the plurality of entities using the location of the one of the plurality of entities.

Abstract: A method includes receiving training data that includes a plurality of sets of text-to-speech (TTS) spoken utterances each associated with a respective language and including TTS utterances of synthetic speech spoken that includes a corresponding reference speech representation paired with a corresponding input text sequence. For each TTS utterance in each set of the TTS spoken training utterances of the received training data, the method includes generating a corresponding TTS encoded textual representation for the corresponding input text sequence, generating a corresponding speech encoding for the corresponding TTS utterance of synthetic speech, generating a shared encoder output, generating a predicted speech representation for the corresponding TTS utterance of synthetic speech, and determining a reconstruction loss. The method also includes training a TTS model based on the reconstruction losses determined for the TTS utterances in each set of the TTS spoken training utterances.

Abstract: A method includes receiving a training dataset that includes one or more spoken training utterances for training an automatic speech recognition (ASR) model. Each spoken training utterance in the training dataset paired with a corresponding transcription and a corresponding target sequence of auxiliary tokens. For each spoken training utterance, the method includes generating a speech recognition hypothesis for a corresponding spoken training utterance, determining a speech recognition loss based on the speech recognition hypothesis and the corresponding transcription, generating a predicted auxiliary token for the corresponding spoken training utterance, and determining an auxiliary task loss based on the predicted auxiliary token and the corresponding target sequence of auxiliary tokens. The method also includes the ASR model jointly on the speech recognition loss and the auxiliary task loss determined for each spoken training utterance.

Abstract: The various implementations described herein include methods and systems for synchronous audio playback. An electronic device can receive an identification of a first device as a common clock device that has a first internal clock being designated as a master clock. The electronic device receives a synchronized audio playback command that includes audio data to be output and a future playback time. In response to receiving the audio data, the device determines a synchronized audio playback time for audio to be output. An optimal time for output can be calculated and transmitted to the server system for future playback time calculations.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for receiving a packet from a client, the packet having header information including a destination Internet Protocol (IP) address, a destination port, a source IP address, and a source port, and wherein the source IP address and source port are associated with the client; selecting a destination virtual machine based on the destination port; modifying the packet by replacing the destination IP address in the header information with an IP address of the selected destination virtual machine; and sending the modified packet to the destination virtual machine.