Nanjing Horizon Robotics Technology Co., Ltd.

Baidu Online Network Technology (beijing) Co., Ltd.

Beijing Horizon Robotics Technology Research and Development Co., Ltd.

Heng Luo

Liang Chen

Chang Huang

Kun Ling

Jianjun Li

Delin Li

Wei Xu

Gaolin Fang

Zhichao Li

Libo Fu

Honghe Tan

Yifeng Geng

Yushu Gao

Disclosed is a method for convolution calculation in a neural network, comprising: reading an input feature map, depthwise convolution kernels and pointwise convolution kernels from a dynamic random access memory (DRAM); performing depthwise convolution calculations and pointwise convolution calculations by depthwise convolution calculation units and pointwise convolution calculation units, according to the input feature map, the depthwise convolution kernels and the pointwise convolution kernels to obtain output feature values of a first predetermined number p of points on all pointwise convolution output channels; storing the output feature values of a first predetermined number p of points on all pointwise convolution output channels into an on-chip memory; and repeating above operation to obtain output feature values of all points on all point wise convolution output channels. Therefore, the storage space for storing intermediate results may be reduced.

Method, electronic device and storage medium for convolution calculation in neural network

Disclosed are a neural network training method, a neural network training device and an electronic device. The neural network training method includes: training a first neural network to be trained by using sample data; determining an indicator parameter of the first neural network in a current training process; determining an update manner corresponding to a preset condition if the indicator parameter meets the preset condition; and updating a parameter of a batch normalization layer in the first neural network based on the update manner. In this way, sparsing of a feature map output by a neural network is implemented, thereby reducing an amount of data to be transmitted and improving computation speed of a chip.

Neural network training method, neural network training apparatus and electronic device

Disclosed is a method for convolution calculation in a neural network, comprising: reading an input feature map, depthwise convolution kernels and pointwise convolution kernels from a dynamitic random access memory (DRAM); performing depthwise convolution calculations and pointwise convolution calculations according to the input feature map, the depthwise convolution kernels and the pointwise convolution kernels to obtain output feature values of a first predetermined number p of points on all pointwise convolution output channels; storing the output feature values of a first predetermined number p of points on all pointwise convolution output channels into an on-chip memory, wherein the first predetermined number p is determined according to at least one of available space in the on-chip memory, a number of the depthwise convolution calculation units, and width, height and channel dimensions of the input feature map; and repeating the above operation obtain output feature values of all points on all pointwise convolution output channels. Therefore, the storage space for storing intermediate results may be reduced.

Method and electronic device for convolution calculation in neural network

Disclosed is an apparatus for performing a convolution operation in a convolutional neural network. The apparatus may comprise a selector for selecting one or more nonzero elements of a weight parameter, a selector for selecting a data item(s) corresponding to selected nonzero elements in input feature data, and a calculator unit for performing an operation. The apparatus may realize the convolution operation in a sparsified convolutional neural network efficiently through the hardware.

Apparatus for performing convolution operations in a convolutional neural network

Embodiments of the present invention disclose a method and an apparatus of establishing an image search relevance prediction model, and an image search method and apparatus. The method of establishing an image search relevance prediction model comprises: training a pre-constructed original deep neural network by using a training sample, wherein the training sample comprises: a query and image data, and the original deep neural network comprises: a representation vector generation network and a relevance calculation network; and using the trained original deep neural network as the image search relevance prediction model. The technical solution of the present invention optimizes the existing image search technology, and has stronger capabilities than the prior art as well as various integrations and variations in terms of semantic matching between a query and an image text, semantic matching between a query and image content, click generalization and the like.

Method and apparatus of establishing image search relevance prediction model, and image search method and apparatus

Growing community of inventors

Beijing, China

Heng Luo