公开(公告)号：US20220156508A1

公开(公告)日：2022-05-19

申请号：US17455201

申请日：2021-11-16

Applicant: QUALCOMM Incorporated

Inventor： Bert MOONS ,  Parham NOORZAD ,  Andrii SKLIAR ,  Christopher LOTT ,  Tijmen Pieter Frederik BLANKEVOORT

IPC: G06K9/62 ,  G06N3/08

Abstract: Various aspects provide methods for a computing device selecting a neural network for a hardware configuration including using an accuracy predictor to select from a search space a neural network including a first plurality of the blockwise knowledge distillation trained search blocks, in which the accuracy predictor is built using search space trained blockwise knowledge distillation search blocks. Aspects may include selecting a second plurality of the blockwise knowledge distillation trained search blocks based on criteria of predicted accuracy using the accuracy predictor for the second plurality of the blockwise knowledge distillation trained search blocks. Aspects may include selecting the neural network based on a search of the blockwise knowledge distillation trained search blocks, initializing the blockwise knowledge distillation trained search blocks of the neural network using weights of the blockwise knowledge distillation trained search blocks, and fine-tuning the neural network using knowledge distillation, to generate a distilled neural network.

公开(公告)号：US20210150306A1

公开(公告)日：2021-05-20

申请号：US17098049

申请日：2020-11-13

Applicant: QUALCOMM Incorporated

Inventor： Jamie Menjay LIN ,  Yang YANG ,  Parham NOORZAD

IPC: G06N3/02 ,  G06F17/15

Abstract: Aspects described herein provide a method of performing phase selective convolution, including: receiving multi-phase pre-activation activation data; partitioning the multi-phase pre-activation data; applying a first activation function to the set of first phase pre-activation data to form a set of first phase activation output; convolving the set of first phase activation output with a first convolution kernel to form a first phase output feature map; negating the set of second phase activation data; applying a second activation function to the negated set of second phase pre-activation data to form a set of second phase activation output; convolving the set of second phase activation output with a second convolution kernel to form a second phase output feature map; negating the second phase output feature map; and training the neural network based on the first phase output feature map and the second phase output feature map.