Baidu USA LLC

Baidu.com Times Technology (beijing) Co. Ltd.

Zhiyu Cheng

Yingze Bao

Le Kang

Hao Tian

Xin Zhou

Baopu Li

Xing Li

Daming Lu

Xi Chen

Jeff Chienyu Wang

Haofeng Kou

Yanwen Fan

Bo He

Jingyu Xin

Dapeng Li

Jingya Xun

Various embodiments for tensor decomposition to compress neural network models are presented. In one or more embodiments, one or more neural network layers from a neural network model are compressed using tensor ring (TR) decomposition. In one or more embodiments, a TR-decomposed neural network utilizes fewer resources and can be more readily deployed on resource-constraint devices. With a smaller model size, during inference time, the compressed model runs faster than its original model and consumes less power. Furthermore, the TR-compressed model can achieve better performance in terms of stability and model accuracy, compared to other tensor decomposition methods.

Tensor ring decomposition for neural networks

Presented herein are systems, methods, and datasets for automatically and precisely generating highlight or summary videos of content. In one or more embodiments, the inputs comprise a text (e.g., an article) of the key event(s) (e.g., a goal, a player action, etc.) in an activity (e.g., a game, a concert, etc.) and a video or videos of the activity. In one or more embodiments, the output is a short video of an event or events in the text, in which the video may include commentary about the highlighted events and/or other audio (e.g., music), which may also be automatically synthesized.

Generating highlight video from video and text inputs

With rapidly evolving technologies and emerging tools, sports-related videos generated online are rapidly increasing. To automate the sports video editing/highlight generation process, a key task is to precisely recognize and locate events-of-interest in videos. Embodiments herein comprise a two-stage paradigm to detect categories of events and when these events happen in videos. In one or more embodiments, multiple action recognition models extract high-level semantic features, and a transformer-based temporal detection module locates target events. These novel approaches achieved state-of-the-art performance in both action spotting and replay grounding. While presented in the context of sports, it shall be noted that the systems and methods herein may be used for videos comprising other content and events.

Transformer-based temporal detection in video

Deep neural networks (DNN) model quantization may be used to reduce storage and computation burdens by decreasing the bit width. Presented herein are novel cursor-based adaptive quantization embodiments. In embodiments, a multiple bits quantization mechanism is formulated as a differentiable architecture search (DAS) process with a continuous cursor that represents a possible quantization bit. In embodiments, the cursor-based DAS adaptively searches for a quantization bit for each layer. The DAS process may be accelerated via an alternative approximate optimization process, which is designed for mixed quantization scheme of a DNN model. In embodiments, a new loss function is used in the search process to simultaneously optimize accuracy and parameter size of the model. In a quantization step, the closest two integers to the cursor may be adopted as the bits to quantize the DNN together to reduce the quantization noise and avoid the local convergence problem.

Cursor-based adaptive quantization for deep neural networks

Presented herein are systems, methods, and datasets for automatically and precisely generating highlight or summary videos of content. For example, in one or more embodiments, videos of sporting events may be digested or condensed into highlights, which will dramatically benefit sports media, broadcasters, video creators or commentators, or other short video creators, in terms of cost reduction, fast, and mass production, and saving tedious engineering hours. Embodiment of the framework may also be used or adapted for use to better promote sports teams, players, and/or games, and produce stories to glorify the spirit of sports or its players. While presented in the context of sports, it shall be noted that the methodologies may be used for videos comprising other content and events.

Automatically and precisely generating highlight videos with artificial intelligence

Embodiments of the present disclosure disclose a method and apparatus for compressing a deep learning model. An embodiment of the method includes: acquiring a to-be-compressed deep learning model; pruning each layer of weights of the to-be-compressed deep learning model in units of channels to obtain a compressed deep learning model; and sending the compressed deep learning model to a terminal device, so that the terminal device stores the compressed deep learning model. By pruning each layer of weights of the deep learning model in units of channels, the parameter redundancy of the deep learning model is effectively reduced, thereby improving the computational speed of the deep learning model and maintaining the model accuracy.

Method and apparatus for compressing deep learning model

A method for information processing for accelerating neural network training. The method includes: acquiring a neural network corresponding to a deep learning task; and performing iterations of iterative training on the neural network based on a training data set. The training data set includes task data corresponding to the deep learning task. The iterative training includes: processing the task data in the training data set using a current neural network, and determining, based on a processing result of the neural network on the task data in a current iterative training, prediction loss of the current iterative training; determining a learning rate and a momentum in the current iterative training; and updating weight parameters of the current neural network by gradient descent based on a preset weight decay, and the learning rate, the momentum, and the prediction loss in the current iterative training. This method achieves efficient and low-cost deep learning-based neural network training.

Method, electronic device and computer readable medium for information processing for accelerating neural network training

Embodiments disclose an artificial intelligence chip and a convolutional neural network applied to the artificial intelligence chip comprising a processor, at least one parallel computing unit, and a pooling computation unit. The method comprises: dividing a convolution task into convolution subtasks and corresponding pooling subtasks; executing convolution subtasks at different parallel computing units, and performing convolution, batch normalization, and non-linear computing operation in a same parallel computing unit; sending an execution result of each parallel computing unit from executing the convolution subtask to the pooling computation unit for executing the corresponding pooling subtask; merging executing results of the pooling computation unit from performing pooling operations on the executing results outputted by respective convolution subtasks to obtain an execution result of the convolution task. This can reduce data transport, such that operations of the convolutional neural network may be accomplished with lower power consumption and less time in an edge device.

Hardware-software co-design for accelerating deep learning inference

Presented herein are embodiments of a training deep learning models. In one or more embodiments, a compact deep learning model comprises fewer layers, which require fewer floating-point operations (FLOPs). Presented herein are also embodiments of a new learning rate function, which can adaptively change the learning rate between two linear functions. In one or more embodiments, combinations of half-precision floating point format training together with larger batch size in the training process may also be employed to aid the training process.

Growing community of inventors

Sunnyvale, CA, United States of America

Zhiyu Cheng