Qualcomm Incorporated

Kofax, Inc.

Intel Corporation

Anthony Sarah

Mauritius A R Schmidtler

Jan Willers Amtrup

Stephen Michael Thompson

Leonard Henry Grokop

Robert Howard Kimball

Sanjiv Nanda

Jeffrey Alexander Levin

Michael-David Nakayoshi Canoy

Roland G Borrey

David Jonathan Julian

Raghuraman Krishnamoorthi

Vidya Narayanan

Jason Frank Hunzinger

Pawan Kumar Baheti

Dexu Lin

Brian Spinar

Somdeb Majumdar

Regan Blythe Towal

Daniel Cummings

Yinyin Liu

Jan Krzys Wegrzyn

Casimir Matthew Wierzynski

Adrienne Milner

Jeffrey Baginsky Gehlhaar

Juan Pablo Munoz

Tristan Webb

Sharath Nittur Sridhar

The present disclosure is related to framework for automatically and efficiently finding machine learning (ML) architectures that are optimized to one or more specified performance metrics and/or hardware platforms. This framework provides ML architectures that are applicable to specified ML domains and are optimized for specified hardware platforms in significantly less time than could be done manually and in less time than existing ML model searching techniques. Furthermore, a user interface is provided that allows a user to search for different ML architectures based on modified search parameters, such as different hardware platform aspects and/or performance metrics. Other embodiments may be described and/or claimed.

Framework for optimization of machine learning architectures

The present disclosure is related to framework for automatically and efficiently finding machine learning (ML) architectures that generalize well across multiple artificial intelligence (AI) and/or ML domains, AI/ML tasks, and datasets. The ML architecture search framework accepts a list of tasks and corresponding datasets as inputs, and may also include relevancy scores/weights for each item in the input. A combined performance metric is generated, where this combined performance metric quantifies the performance of the ML architecture across all the specified AI/ML domains, AI/ML tasks, and datasets. The system then performs a multi-objective ML architecture search with the combined performance metric, along with hardware-specific performance metrics as the objectives. Other embodiments may be described and/or claimed.

Hardware-aware machine learning model search mechanisms

An apparatus of operating a computational network is configured to determine a low-rank approximation for one or more layers of the computational network based at least in part on a set of residual targets. A set of candidate rank vectors corresponding to the set of residual targets may be determined. Each of the candidate rank vectors may be evaluated using an objective function. A candidate rank vector may be selected and used to determine the low rank approximation. The computational network may be compressed based on the low-rank approximation. In turn the computational network may be operated using the one or more compressed layers.

Determining layer ranks for compression of deep networks

According to one embodiment, a computer-implemented method for cleaning up a data set having a possible incorrect label includes: selecting a plurality of training documents; estimating a quality of an organization of a plurality of categories; and determining whether the quality of the organization is greater than a predetermined quality threshold. Corresponding system and computer program product embodiments are also presented. Other aspects and advantages of the present invention will become apparent from the following detailed description, which, when taken in conjunction with the drawings, illustrate by way of example the principles of the invention.

Systems and methods for organizing data sets

A method of online training of a classifier includes determining a distance from one or more feature vectors of an object to a first predetermined decision boundary established during off-line training for the classifier. The method also includes updating a decision rule as a function of the distance. The method further includes classifying a future example based on the updated decision rule.

Online training for object recognition system

A method of generating executable code for a target platform in a neural network includes receiving a spiking neural network description. The method also includes receiving platform-specific instructions for one or more target platforms. Further, the method includes, generating executable code for the target platform(s) based on the platform-specific instructions and the network description.

Compiling network descriptions to multiple platforms

According to one embodiment, a computer-implemented method for confirming/rejecting a most relevant example includes: generating a binary decision model by training a binary classifier using a plurality of training documents; classifying one or more test documents into one of a plurality of categories using the binary decision model, wherein the one or more test documents lack a user-defined category label; selecting a most relevant example of the classified test documents from among the classified test documents; displaying, using a display of the computer, the most relevant example of the classified test documents to a user; receiving, via the computer and from the user, a confirmation or a negation of a classification label of the most relevant example of the classified test documents; and storing the confirmation or the negation of the classification label of the most relevant example of the classified test documents to a memory of the computer.

Certain aspects of the present disclosure support efficient implementation of common neuron models. In an aspect, a first memory layout can be allocated for parameters and state variables of instances of a first neuron model, and a second memory layout different from the first memory layout can be allocated for parameters and state variables of instances of a second neuron model having a different complexity than the first neuron model.

Method and apparatus for efficient implementation of common neuron models

A method for dynamically setting a neuron value processes a data structure including a set of parameters for a neuron model and determines a number of segments defined in the set of parameters. The method also includes determining a number of neuron types defined in the set of parameters and determining at least one boundary for a first segment.

Defining dynamics of multiple neurons

Methods and apparatus are provided for training a neural device having an artificial nervous system by modulating at least one training parameter during the training. One example method for training a neural device having an artificial nervous system generally includes observing the neural device in a training environment and modulating at least one training parameter based at least in part on the observing. For example, the training apparatus described herein may modify the neural device's internal learning mechanisms (e.g., spike rate, learning rate, neuromodulators, sensor sensitivity, etc.) and/or the training environment's stimuli (e.g., move a flame closer to the device, make the scene darker, etc.). In this manner, the speed with which the neural device is trained (i.e., the training rate) may be significantly increased compared to conventional neural device training systems.

Methods and apparatus for modulating the training of a neural device

A method for dynamically modifying synaptic delays in a neural network includes initializing a delay parameter and operating the neural network. The method further includes dynamically updating the delay parameter based on a program which is based on a statement including the delay parameter.

Dynamically assigning and examining synaptic delay

Components, methods, and apparatuses are provided that may be used to obtain a plurality of time-referenced features extracted from signals received from one or more inertial sensors of a mobile device, wherein the time-referenced features are provided according to heterogeneous time scales. The time-referenced features are combined over a decision window to infer an activity of a user co-located with the mobile device.

Methods, devices, and apparatuses for activity classification using temporal scaling of time-referenced features

Information about a context of a mobile device is inferred using measurements from light detectors of the mobile device. The light detectors may be located on multiple different sides of the mobile device. In some embodiments, measurements from one or more cameras may be used to infer context.

Method and apparatus for characterizing context of a mobile device

A method is provided for organizing data sets. In use, an automatic decision system is created or updated for determining whether data elements fit a predefined organization or not, where the decision system is based on a set of preorganized data elements. A plurality of data elements is organized using the decision system. At least one organized data element is selected for output to a user based on a score or confidence from the decision system for the at least one organized data element. Additionally, at least a portion of the at least one organized data element is output to the user. A response is received from the user comprising at least one of a confirmation, modification, and a negation of the organization of the at least one organized data element. The automatic decision system is recreated or updated based on the user response. Other embodiments are also presented.

A method for improving neural dynamics includes obtaining prototypical neuron dynamics. The method also includes modifying parameters of a neuron model so that the neuron model matches the prototypical neuron dynamics. The neuron dynamics comprise membrane voltages and/or spike timing.

Growing community of inventors

San Diego, CA, United States of America

Anthony Sarah