One minute quick view！
Tesla develops Dojo neural network training computer, which is a performance beast
L there are more than 400 vulnerabilities in Qualcomm snapdragon chips, affecting more than 40% of the world’s models
Strategic cooperation: Ctrip‘s core supply links to Jingdong
Intel 11th generation core achieves “cross generation” promotion: higher frequency under low voltage
NVIDIA’s acquisition of arm, worth about $44 billion, could be completed as soon as the end of summer
L a few lines of code can effectively create data sets, Google open source tfrecorder
CMU researchers explore sound to help improve Robot Perception
L Google open source chromeOS.dev It’s easier to build applications on Chrome OS
L ACL 2020 | search better transformer structure based on different hardware
L ECCV 2020 spotlight | fine grained region similarity self monitoring in image location
L eccv2020 significantly improves segmentation prediction, ETH open source video target segmentation based on scenario storage network
1、Tesla develops Dojo neural network training computer, which is a performance beast
On August 16, Tesla CEO musk said on the social network recently that the company is developing a neural network training computer called Dojo to process a large amount of video data. According to musk, dojo is like a real performance beast. At last year’s autonomy day, musk said that Dojo’s goal is to be able to receive a large amount of data and conduct training at the video level, and to use Dojo programs or Dojo computers to conduct unsupervised mass training on a large number of videos.
2、There are more than 400 vulnerabilities in Qualcomm snapdragon chips, affecting more than 40% of the world’s models
Check point, a network security provider, said that in a study codenamed Achilles, the company conducted an extensive security assessment of Qualcomm snapdragon’s digital signal processing (DSP) chips. The results show that there are more than 400 vulnerable code segments in the chip. This means that more than 40% of devices in the global market (involving Android phones such as Google, Samsung, LG, Xiaomi and Yijia) will be affected by the vulnerability and face the risk of being hacked. DSP chip is an auxiliary chip in mobile phones, which is mainly responsible for processing audio, video and image data. It appears in most modern mobile phones and is provided with Qualcomm’s snapdragon processor.
3、Strategic cooperation: Ctrip’s core supply links to Jingdong
On August 16, Jingdong group and Ctrip Group officially signed a strategic cooperation agreement. According to the cooperation agreement, Ctrip’s core product supply chain will be connected to Jingdong platform, and the two sides will carry out all-round cooperation in user flow, channel resources, cross-border marketing, business travel development, e-commerce cooperation, etc. After the cooperation, Ctrip will provide real-time product inventory and competitive product prices for Jingdong, and Jingdong will access the core product supply chain of Ctrip, and open the user flow of Jingdong platform to Ctrip, so as to provide comprehensive support for the product supply chain of Ctrip in terms of daily operation and precision marketing.
4、Intel 11th generation core achieves “cross generation” promotion: higher frequency under low voltage
Intel officially announced tiger Lake SOC, which adopts willow Cove architecture, and officially said it will provide performance improvement beyond intergenerational CPU. According to Intel’s official documents, under the willow Cove architecture, the processor can achieve higher frequency at lower voltage.
5、NVIDIA’s acquisition of arm, worth about $44 billion, could be completed as soon as the end of summer
NVIDIA and arm have entered into exclusive negotiations and are expected to reach a deal by the end of the summer, people familiar with the matter said. According to foreign media, in April this year, arm was listed for sale by its parent Softbank for the first time when Goldman Sachs, the US investment bank, was hired to look for potential buyers. Goldman contacted apple in April, but apple did not intend to participate in the bidding, because arm’s licensing business is not very consistent with Apple’s business model of combining software and hardware. Moreover, if Apple buys the chip technology licensing company, which supplies many competitors, it may also cause regulatory concerns.
6、A few lines of code can effectively create data sets, Google open source tfrecorder
Google recently opened the tensorflow recorder (tfrecorder) project to simplify the creation process of tfrecord. Tfrecord is a binary file format, which is relatively efficient in data processing. However, it is troublesome to convert other data into tfrecord. It is usually necessary to write a data pipeline to parse structured data, load images from storage, and then serialize the results into tfrecord format. The open source tfrecorder can write tfrecords directly from pandas dataframe or CSV format, without writing complex code.
7、CMU researchers explore sound to help improve Robot Perception
A new experiment by Carnegie Mellon University’s research team suggests moving objects in metal pallets using retreating robotics’s Sawyer to make them feel the sound they make as they roll around, slide and hit the side. A total of 60 objects – including tools, blocks, tennis and an apple – recorded and categorized 15000 “interactions.”. The team named the robot “tilt bot”, which can recognize objects with a success rate of 76%, and even identify relatively small material differences between metal screwdrivers and spanners. By using sound data, the robot is usually able to correctly determine the material composition of the object.
8. Google open source chromeOS.dev It’s easier to build applications on Chrome OS
Google recently launched the Chrome OS Developer Center（ chromeOS.dev ）The website contains many technical resources, tutorials, code examples and fresh information. Its purpose is to “help developers make greater use of the functions on the platform”, understand Chrome OS and build Chrome OS applications more easily. chromeOS.dev The main target groups are web, Android and Linux developers, as well as designers, product managers and business leaders. At present, the website only provides English and Spanish, and will support more languages in the future.
1、 ACL 2020 | search better transformer structure based on different hardware
In the past, a lot of work has been done to simplify the transformer architecture, but they have not considered the influence of different hardware on the model architecture. In this paper, we first propose the method of using network architecture search (NAS) to search the best transformer architecture for different hardware. Experiments show that the transformer structure is smaller and faster than other models under different hardware, and does not damage the effect.
Link to the paper: https://arxiv.org/abs/2005.14187
2、 ECCV 2020 spotlight | fine grained region similarity self-monitoring for image location
This paper introduces a paper published in ECCV 2020, which is honored to be included as spotlight presentation. We propose an effective solution to the problem of weak supervision in large-scale image location, aiming at fully mining the difficult samples in representation learning through self supervised learning, and further fine-grained image level supervision into regional level supervision, so as to better model the complex relationship between images and regions. The model trained by this algorithm has strong robustness and generalization, which is verified on several image location data sets, [email protected] The accuracy is 5.7% higher than that of the most advanced technology, and the code and model have been published.
Link to the paper: https://arxiv.org/abs/2006.03926
3、Eccv2020 significantly improves segmentation prediction, ETH open source video object segmentation based on scene map storage network
This paper focuses on solving a basic problem in the field of video object segmentation: making the segmentation model adapt to the appearance changes of specific video and online objects effectively. A simple and fast new graph storage mechanism is proposed, which significantly improves segmentation prediction. In addition, the framework generated by graph storage network can also be extended to one shot and zero shot video object segmentation tasks.
Link to the paper: https://arxiv.org/pdf/2007.07…