Nanjing Horizon Robotics Technology Co., Ltd.

Tsinghua University

Omron Corporation

Nec Laboratories America, Inc.

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

Beijing Horizon Information Technology Co., Ltd

Other

Facebook, Inc.

Meta Platforms, Inc.

Beijing Baidu Netcom Science and Technology Co., Ltd.

Chang Huang

Liang Chen

Kun Ling

Jianjun Li

Delin Li

Shihong Lao

Haizhou Ai

Qian Zhang

Feng Zhou

Heng Luo

Shenghuo Zhu

Kai Hung Yu

Takayoshi Yamashita

Zhenjiang Wang

Guoli Wang

Bo Wu

Yuan Yuan Li

Guang Cheng Li

Tian Xia

Emanuel Alexandre Strauss

Kun Hsien Lin

Honghe Tan

Yi Zeng

Chaoxu Guo

Yonghao Xu

Wenhui Cao

Yangqing Jia

Katherine Ruolin Yu

Yukang Chen

Akshita Rajendra Jain

Zhuoran Zhao

Haoqian He

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 convolution processing engine and a control method thereof, and a convolutional neural network accelerator comprising the convolution processing engine. The convolution processing engine comprises at least two cache memories connected in series and an operational circuit. The convolution processing engine can realize an efficient convolution operation with lower complexity and power consumption.

Convolution processing engine and control method, and corresponding convolutional neural network accelerator

Disclosed are a data compression method, a computer-readable storage medium, and an electronic device. The method includes: converting each data in a to-be-compressed data set into binary data in a preset format; determining a to-be-compressed bit and a significant bit for the each data in the to-be-compressed data set based on a sequence of all bits of the binary data; determining a compression bit width corresponding to the to-be-compressed data set based on bit widths of the significant bits; compressing the each data in the to-be-compressed data set based on the compression bit width, to obtain a compressed data set; and generating attribute information of the compressed data set. According to the present disclosure, the significant bit can be determined based on the sequence of all bits without adjusting orders of the bits of the binary data, thereby simplifying a data compression process and improving efficiency of data compression.

Data compression method and apparatus, computer-readable storage medium, 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

A method and an apparatus for adapting feature data in a convolutional neural network. The method includes selecting a plurality of consecutive layers; determining an expected number of subdata blocks and a layout position, width and height of each subdata block in an output feature data of a last layer; determining, for each current layer, a layout position, width, and height of each subdata block of an input feature data for the current layer according to the layout position, width, and height of each subdata block of the output feature data for the current layer; determining an actual position of each subdata block of the input feature data for a first layer in the input feature data for the first layer; and obtaining the expected number of subdata blocks of the input feature data for the first layer according to the actual position, width and height of each subdata block of the input feature data for the first layer.

Method and apparatus for adapting feature data in a convolutional neural network

An online system receives digital content and determines whether the digital content includes malicious content, such as obfuscated text, before presenting the digital content to a user. To determine whether the digital content contains malicious content, the online system renders the digital content. The online system performs optical character recognition on the content. The online system uses an obfuscation machine learning model to identify obfuscated text. The online system may deobfuscate the obfuscated text. The online system may prevent presentation of the digital content in response to detecting obfuscated text.

Identification and deobfuscation of obfuscated text in digital content

Disclosed are a method and an apparatus for adapting parameters of a neural network. The method includes selecting one or more dimensions for a weight parameter of each of at least one layer of the neural network, determining a dimension value and a corresponding target value in each dimension of the weight parameter, and padding the weight parameter in a case where the dimension value in at least one dimension of the weight parameter is less than the corresponding target value, the dimension value in each dimension of the weight parameter after the padding being equal to the corresponding target value.

Method and apparatus for adapting parameters of 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

A method for data management is provided. The method comprises: storing the plurality of items in a contiguous space within the memory, executing an instruction containing an address and a size that together identify the contiguous space to transmit the plurality of items from the main memory to a random-access memory (RAM) on a chip, and the chip includes a computing unit comprising a plurality of multipliers; and instructing the computing unit on the chip to: retrieve multiple of the plurality of items from the RAM; and perform a plurality of parallel operations using the plurality of multipliers with the multiple items to yield output data.

Systems and methods for data management

Disclosed are a method for updating a neural network and an electronic device. The method includes: inputting a first image set having tag information into a first depth neural network, and determining a cross entropy loss value of the first image set by using the first depth neural network; inputting a second image set having no tag information separately into the first depth neural network and a second depth neural network, and determining a consistency loss value of the second image set, the first depth neural network and the second depth neural network having the same network structure; updating parameters of the first depth neural network based on the cross entropy loss value and the consistency loss value; and updating parameters of the second depth neural network based on the updated parameters of the first depth neural network.

Method for updating neural network and electronic device

Disclosed are a method for generating a neural network, an apparatus thereof, and an electronic device. The method includes: obtaining an optimal neural network and a worst neural network from a neural network framework by using an evolutionary algorithm; obtaining an optimized neural network from the optimal neural network by using a reinforcement learning algorithm; updating the neural network framework by adding the optimized neural network into the neural network framework and deleting the worst neural network from the neural network framework; and determining an ultimately generated neural network from the updated neural network framework. In this way, a neural network is optimized and updated from a neural network framework by combining the evolutionary algorithm and the reinforcement learning algorithm, thereby automatically generating a neural network structure rapidly and stably.

Method for generating neural network and electronic device

An image recognition method includes: determining a first feature map of the current frame image by using a convolutional neural network based on a type of a current frame image; determining a second feature map of a key frame image before the current frame image; performing feature alignment on the first feature map and the second feature map to obtain a first aligned feature map; fusing the first feature map and the first aligned feature map to obtain a first fused feature map; and recognizing content in the current frame image based on the first fused feature map.

Image recognition method, electronic apparatus and readable storage medium

A method and apparatus are disclosed to perform the circular addressing to emulate a virtually unlimited memory space despite the fixed capacity of a physical memory by readdressing the portion of the data that exceeds the pre-defined length of the circular addressing region to another pre-defined address in the circular addressing region. Data segments in a data sample can be loaded and computed with recalculated circular addresses for different applications.

Methods and apparatus for using circular addressing in convolutional operation

In one embodiment, a method includes receiving an text query, identifying and ranking multiple content objects corresponding to character strings from the text query. A computing system determines the ranking based on an analysis of the association between the terms in the content objects and the n-grams in the query using term frequency-inverse document frequency (TF-IDF) scores of the terms. The computing system accesses a query-term index to determine the TF-IDF scores for the terms, wherein the query-term index includes multiple entries corresponding to multiple terms extracted from multiple prior queries. Each entry in the query-term index includes a query term, one or more related terms extracted from content objects of the online social network matching the prior query term, and a term frequency-inverse document frequency (TF-IDF) score for the respective related terms.

Determining related query terms through query-post associations on online social networks

Growing community of inventors

Beijing, China

Chang Huang