Apple Inc.

Other

Umesh Suresh Vaishampayan

Gaurav Kapoor

Andrew H Vyrros

Abhradeep Guha Thakurta

Abhishek Bhowmick

Kit-Man Wan

Vipul Ved Prakash

Julien F Freudiger

James Michael Magee

Phillip Stanley-Marbell

Yannick L Sierra

Alexander Barraclough Brown

Xiaojin Shi

Marco Zuliani

Matthew John Byom

Steven J Falkenburg

Arnaud Legendre

Steven Duplinsky

Cecile Marie Foret

Vivek Rangarajan Sridhar

Doug Davidson

Julien Freudinger

Kevin W Decker

Michael Brouwer

Russell Alexader Blaine

Kartik R Venkatraman

Conrad A Shultz

Francesco Rossi

John C Hurley

Ryan M Rogers

Daniel Heller

Keaton F Mowery

Mateusz Rajca

Steve Falkenburg

Etienne Belanger

Daniel Allen Chimene

Michael R Siracusa

Derek R Kumar

Nils Nieuwejaar

Joao Pedro Forjaz De Lacerda

Alexey Marinichev

Matthew R Salesi

Joseph Sokol, Jr

Kevin James Van Vechten

Guy G Sotomayor

Gavin Barraclough

Damien Pascal Sorresso

Christopher John Sarcone

Benjamin C Trumbull

Tejas Arun Bahulkar

Timothy Paul Marsland

Adam C Swift

Anthony Graham Sumpter

Ananthakrishna Ramesh

Christopher G Peak

Benjamin H Nham

Albert Antony

Christophe Dumez

Shachar Katz

Harshavardhan P Gopalakrishnan

William E Damon, Iii

Yosen Lin

Christophe P Dumez

Mateusz K Rajca

Daniel S Heller

A device implementing a system to execute machine learning models from memory includes at least one processor configured to receive a request to provide an input to one or more machine learning (ML) models arranged into a graph of connected layers, the one or more ML models stored in the first type of memory. The at least one processor is further configured to divide the graph of connected layers into a plurality of segments such that at least two of the plurality of segments concurrently fits within allocated space of the second type of memory. The at least one processor is further configured to cause the input to be processed through the first segment of the plurality of segments using the second type of memory while a second segment of the plurality of segments is concurrently loaded from the first type of memory into the second type of memory.

Execution of segmented machine learning models

Embodiments described herein provide a technique to crowdsource labeling of training data for a machine learning model while maintaining the privacy of the data provided by crowdsourcing participants. Client devices can be used to generate proposed labels for a unit of data to be used in a training dataset. One or more privacy mechanisms are used to protect user data when transmitting the data to a server. The server can aggregate the proposed labels and use the most frequently proposed labels for an element as the label for the element when generating training data for the machine learning model. The machine learning model is then trained using the crowdsourced labels to improve the accuracy of the model.

Distributed labeling for supervised learning

The subject technology provides receiving a neural network (NN) model to be executed on a target platform, the NN model including multiple layers that include operations and some of the operations being executable on multiple processors of the target platform. The subject technology further sorts the operations from the multiple layers in a particular order based at least in part on grouping the operations that are executable by a particular processor of the multiple processors. The subject technology determines, based at least in part on a cost of transferring the operations between the multiple processors, an assignment of one of the multiple processors for each of the sorted operations of each of the layers in a manner that minimizes a total cost of executing the operations. Further, for each layer of the NN model, the subject technology includes an annotation to indicate the processor assigned for each of the operations.

Compiling models for dedicated hardware

In some implementations, a user device can maintain a multi-device context store. For example, the user device can receive device and/or user context information from multiple devices and store the context information in a local data store. The user device can collect local device and/or user context information and store the context information in the local context store. The user device can receive user/device context queries from client processes and send the client processes user/device context information from multiple devices in response to the queries.

Multi-device context store

The subject technology runs a compiled neural network (NN) model on a particular processor with multiple priority queues for executing different processes, the compiled NN model being assigned to a particular priority queue, and the compiled NN model includes context switch instructions that were previously inserted into a neural network (NN) model from which the compiled NN model was compiled. The subject technology determines that a particular context switch instruction has been executed by the particular processor. The subject technology determines that a different process is waiting to be executed, the different process being assigned to a different priority queue and the different process being a higher priority process than the running compiled NN model. In response to executing the particular context switch instruction, the subject technology performs a context switch to the different process assigned to the different priority queue when the different process is waiting to be executed.

Disclosed are systems, methods, and non-transitory computer-readable storage media for efficiently monitoring the operating context of a computing device. In some implementations, the context daemon and/or the context client can be terminated to conserve system resources. For example, if the context daemon and/or the context client are idle, they can be shutdown to conserve battery power or free other system resources (e.g., memory). When an event occurs (e.g., a change in current context) that requires the context daemon and/or the context client to be running, the context daemon and/or the context client can be restarted to handle the event. Thus, system resources can be conserved while still providing relevant context information collection and callback notification features.

Efficient context monitoring

Embodiments described herein ensure differential privacy when transmitting data to a server that estimates a frequency of such data amongst a set of client devices. The differential privacy mechanism may provide a predictable degree of variance for frequency estimations of data. The system may use a multibit histogram model or Hadamard multibit model for the differential privacy mechanism, both of which provide a predictable degree of accuracy of frequency estimations while still providing mathematically provable levels of privacy.

Differential privacy using a multibit histogram

In an exemplary process for remote execution of machine-learned models, one or more signals from a second electronic device is detected by a first electronic device. The second electronic device includes a machine-learned model associated with an application implemented on the first electronic device. Based on the one or more signals, a communication connection is established with the second electronic device and a proxy to the machine-learned model is generated. Input data is obtained via a sensor of the first electronic device. A representation of the input data is sent to the second electronic device via the proxy and the established communication connection. The representation of the input data is processed through the machine-learned model to generate an output. A result derived from the output is received via the communication connection and a representation of the result is outputted.

Remote execution of machine-learned models

Embodiments described herein provide a privacy mechanism to protect user data when transmitting the data to a server that estimates a frequency of such data amongst a set of client devices. In one embodiment, a differential privacy mechanism is implemented using a count-mean-sketch technique that can reduce resource requirements required to enable privacy while providing provable guarantees regarding privacy and utility. For instance, the mechanism can provide the ability to tailor utility (e.g. accuracy of estimations) against the resource requirements (e.g. transmission bandwidth and computation complexity).

User experience using privatized crowdsourced data

Embodiments described herein provide techniques to encode sequential data in a privacy preserving manner before the data is sent to a sequence learning server. The server can then determine aggregate trends within an overall set of users, without having any specific knowledge about the contributions of individual users. The server can be used to learn new words generated by user client devices in a crowdsourced manner while maintaining local differential privacy of client devices. The server can also learn other sequential data including typed, autocorrected, revised text sequences, sequences of application launches, sequences of purchases on an application store, or other sequences of activities that can be performed on an electronic device.

Privatized apriori algorithm for sequential data discovery

One embodiment provides a system that implements a 1-bit protocol for differential privacy for a set of client devices that transmit information to a server. Implementations may leverage specialized instruction sets or engines built into the hardware or firmware of a client device to improve the efficiency of the protocol. For example, a client device may utilize these cryptographic functions to randomize information sent to the server. In one embodiment, the client device may use cryptographic functions such as hashes including SHA or block ciphers including AES to provide an efficient mechanism for implementing differential privacy.

Efficient implementation for differential privacy using cryptographic functions

Growing community of inventors

Santa Clara, CA, United States of America

Umesh Suresh Vaishampayan