公开(公告)号：US10510146B2

公开(公告)日：2019-12-17

申请号：US15422395

申请日：2017-02-01

Applicant: QUALCOMM Incorporated

Inventor： Sachin Subhash Talathi

IPC: G06K9/00 ,  G06T7/10 ,  G06K9/46 ,  G06T3/40 ,  G06T5/00 ,  G06T9/00

Abstract: A method for processing an input in an artificial neural network (ANN) includes receiving, at an operator layer of a set of operator layers, a first feature value based on the input from a decoder convolutional layer of a decoder. The operator layer also receives a second feature value based on the input from an encoder convolutional layer of a encoder. The method also includes determining, at the operator layer, a third feature value based on the input by performing an element-wise operation with the first feature value based on the input and the second feature value based on the input. The method transmits, from the operator layer, the third feature value based on the input to an encoder layer that is subsequent to the encoder convolutional layer. The method generates an output based on the third feature value based on the input.

公开(公告)号：US10474949B2

公开(公告)日：2019-11-12

申请号：US14528890

申请日：2014-10-30

Applicant: QUALCOMM Incorporated

Inventor： Somdeb Majumdar ,  Regan Blythe Towal ,  Sachin Subhash Talathi ,  David Jonathan Julian ,  Venkata Sreekanta Reddy Annapureddy

IPC: G06N3/08 ,  G06N3/04 ,  G06N7/00

Abstract: A method for classifying an object includes applying multiple confidence values to multiple objects. The method also includes determining a metric based on the multiple confidence values. The method further includes determining a classification of a first object from the multiple objects based on a knowledge-graph when the metric is above a threshold.

公开(公告)号：US10332028B2

公开(公告)日：2019-06-25

申请号：US14863410

申请日：2015-09-23

Applicant: QUALCOMM Incorporated

Inventor： Sachin Subhash Talathi ,  Aniket Vartak

IPC: G06N3/02 ,  G06N20/00 ,  G06N3/08 ,  G06K9/62 ,  G06K9/46 ,  G06N3/04

Abstract: A method for improving performance of a trained machine learning model includes adding a second classifier with a second objective function to a first classifier with a first objective function. Rather than minimizing a function of errors for the first classifier, the second objective function is used to directly reduce the number errors of the first classifier.

公开(公告)号：US10043112B2

公开(公告)日：2018-08-07

申请号：US14528935

申请日：2014-10-30

Applicant: QUALCOMM Incorporated

Inventor： Sachin Subhash Talathi ,  David Jonathan Julian

IPC: G06F17/30 ,  G06K9/62 ,  G06K9/66 ,  G06N3/04 ,  G06N99/00 ,  G06K9/46

Abstract: A method for image processing includes determining features of multiple stored images from a pre-trained deep convolutional network. The method also includes clustering each image of the multiple stored images based on the determined features.

公开(公告)号：US10339447B2

公开(公告)日：2019-07-02

申请号：US14449101

申请日：2014-07-31

Applicant: QUALCOMM Incorporated

Inventor： Sachin Subhash Talathi ,  David Jonathan Julian ,  Venkata Sreekanta Reddy Annapureddy

IPC: G06N3/08 ,  G06N3/04

Abstract: A method for selecting a reduced number of model neurons in a neural network includes generating a first sparse set of non-zero decoding vectors. Each of the decoding vector is associated with a synapse between a first neuron layer and a second neuron layer. The method further includes implementing the neural network only with selected model neurons in the first neuron layer associated with the non-zero decoding vectors.

公开(公告)号：US10275719B2

公开(公告)日：2019-04-30

申请号：US14848296

申请日：2015-09-08

Applicant: QUALCOMM Incorporated

Inventor： Sachin Subhash Talathi ,  David Jonathan Julian

IPC: G06N20/00 ,  G06N3/08 ,  G06N7/00

Abstract: Hyper-parameters are selected for training a deep convolutional network by selecting a number of network architectures as part of a database. Each of the network architectures includes one or more local logistic regression layer and is trained to generate a corresponding validation error that is stored in the database. A threshold error for identifying a good set of network architectures and a bad set of network architectures may be estimated based on validation errors in the database. The method also includes choosing a next potential hyper-parameter, corresponding to a next network architecture, based on a metric that is a function of the good set of network architectures. The method further includes selecting a network architecture, from among next network architectures, with a lowest validation error.