.JPG)
Key news (1016～1022)
Facebook Electrocatalyst renewable energy
Facebook teamed up with CMU to use AI to find electrocatalysts that convert energy to fuel
Facebook has teamed up with Carnegie Mellon University (CMU) to launch an Open Catalyst Project a few days ago, which will use AI to accelerate the identification of suitable electrocatalysts to store and use renewable energy on a large scale. In response to climate change, many countries have gradually developed renewable energy sources, such as solar or wind energy, to generate intermittent electricity. When these intermittent electric energy is greater than the consumption, it will be temporarily stored, and will be released to the grid when the consumption is greater than the output.
Therefore, another method of renewable energy storage is derived, which is to convert excess wind or solar energy into fuel. However, this requires expensive electrocatalysts (such as platinum) to carry out chemical reactions. If large-scale applications are to be used, a lower-cost electrocatalyst must be found. For this reason, Facebook and CMU initiated this project to use AI to find a new structure of electrocatalyst to compensate for the high computational cost of quantum simulation. At the same time, the team also open sourced the project’s data set, including 1.2 million molecular relaxations and 250 million DFT calculations. (Full text)
Deep heart Quantum mechanics Fermi
DeepMind uses AI to simplify quantum equations and accelerate the development of new materials
DeepMind recently open-sourced a set of Fermi sub-class neural network FermiNet, which can simplify the original complex wave function calculation process and find out the probability of electron position distribution in the quantum state to simulate the influence of chemical bonds. DeepMind pointed out that FermiNet is currently the most accurate model. After open source, the community in related fields can be used to simulate the prototype of a compound or new material and then develop it.
DeepMind said that it is difficult to locate electrons in a quantum state, and this FermiNet can accurately predict the position of electrons in a quantum system. FermiNet will randomly select the electronic arrangement, evaluate the energy of each arrangement, and add them up and minimize them to obtain an approximate value of the true energy. This method, also known as Monte Carlo, is like a gambler who rolls the dice over and over again; in order to obtain a value that is more approximate to the real energy, you can “roll one more dice at any time.” Using this principle, FermiNet uses the original calculation of particles The wave function of the position probability is used to generate the particle position samples and the training data to train itself; that is, in addition to the nucleus position, FermiNet does not require any external training data.
Later, DeepMind also applied this model to an AI system, which can accurately predict the motion trajectory of glass molecules from liquid to solid. Following the publication of the paper at the beginning of the year, the team is now open source FermiNet code to promote the research and development of related communities. (Full text)
loan automation Borrow DocAI
Aiming at financial mortgages, Google launches Lending DocAI to accelerate document review
Google launched Lending DocAI, a cloud-based solution that specifically locks in the mortgage business. It uses Google’s optical character recognition (OCR) and natural language processing (NLP) technologies to automatically process document review tasks by automatically capturing key document data, like It is the income and asset documents of the borrower to speed up the lending process. Google claims that the service can provide industry-leading accuracy.
Google pointed out that mortgage lending is one of the most time-consuming business, because financial institutions have to manually browse hundreds of pages of documents to approve loans. To this end, the team developed Lending DocAI to automate the document review process by capturing document data. Currently, Google DocAI is a preview version, and the official version launch time is undetermined. (Full text)
Old city simulation Computer vision 3D modeling
Google open source AI old city simulation tool set, show you the style of a century-old city
Are you curious about how your home city looked like a hundred years ago? Google AI recently open-sourced a browser tool set rǝ, which allows users to query the 3D appearance changes of a place from 1800 to 2000 on a map on a web page. Google pointed out that the tool runs on its own cloud and Kubernetes and is highly scalable. It can reconstruct old stone streetscapes and buildings in 3D through historical maps and photos.
This tool set consists of three parts. The first is the Warper mass outsourcing platform, which allows users to upload historical maps and photos, and can also correct the geographic coordinates of the old city. Next is the timeline. The user can drag and drop on the timeline part, and the screen will change explicitly with time. Finally, the 3D experience platform. The team uses deep learning models such as regional convolutional model (RCNN) and semantic segmentation model DeepLab to identify different types of objects. Finally, these objects are formed into a 3D grid to form the final appearance of the building. However, the platform is still unfinished, and Google also welcomes masters from all parties to contribute historical atlas. (Full text)
Baoya Smart shelf Precision marketing
Smart Shelf was unveiled in Baoya Minsheng Store, moving towards precise marketing based on the perception of the flow of people
Another case of domestic smart retail! Baoya recently introduced smart shelves in the Taipei Minsheng Store, combining advertising broadcast and human flow perception technology to collect consumer behavior data in front of the shelves to help brand owners understand the benefits of product promotion.
The shelf in the daily necessities section of this Baoya store concisely displays the products, the display on the shelf displays dynamic prices and advertising content; there are also a screen and a set of lenses above the shelf, which usually display product promotion information, when the consumer is detected close , The screen will automatically play brand advertisements related to the products on the shelf. In addition, it can analyze the flow of people, count how many people are close to the shelf because of interest, and even identify the gender of consumers, determine the possible age distribution, and use it as a product benefit analysis . (Full text)
arXiv Code Machine learning
Facebook open source NLP model, 100 languages can be translated without relying on English
Facebook recently open-sourced the M2M-100 natural language processing model, which claims to be the world’s first AI model that can directly translate two languages without an English intermediary. Generally, machine translation will be centered on English. For example, when translating Chinese and French, it will first translate from French to English and then from English to Chinese, but this translation method will cause loss of original meaning or translation errors.
And Facebook’s M2M-100 uses new exploration technology to obtain translation data and build a many-to-many data set covering 7.5 billion sentences in 100 languages and 2,200 translation directions (such as Chinese to French). Finally, they used multiple techniques to train this single model with 15 billion parameters. After the BLEU benchmark test, the model is as good as traditional bilingual translation, and it is 10 points higher than the current translation model of Facebook. (Full text)
Microsoft AI Illustrated Computer vision
Word, Outlook and PowerPoint can all be said by AI with pictures
Microsoft announced a few days ago that it will integrate the AI graphics function into the Windows and Mac versions of Word and Outlook, as well as the Windows, Mac and Web versions of PowerPoint to automatically generate graphics for the images in the file.
In the past, graphic systems were usually trained with data sets containing pictures and text descriptions, but such systems could not identify objects that did not appear in the data set. Therefore, Microsoft pre-trained a large AI model with a rich data set, matched with pictures. Text labels, as visual vocabulary, are then trained with graphic data sets to allow the visual vocabulary model to learn how to compose sentences, so that the new graphic system can identify and describe pictures more accurately, and perform better than humans. (Full text)
IT system detection Current service Watson AI
IT maintenance helper! ServiceNow teamed up with IBM to create an IT system automatic diagnosis solution
ServiceNow announced on its blog that it will integrate IBM’s Watson AI technology with its own IT Service Management System (ITSM) and IT Operation Management System (ITOM) to provide solutions for automatic problem detection and repair. release.
ServiceNow’s ITSM system allows IT personnel to perform scalable IT service management, including incident management and problem management, on a single cloud platform, while ITOM allows IT personnel to explore and view, and deploy all IT in the data center across the cloud and on-site Resources. The joint solution created with IBM allows users to fully understand the operational footprint and respond to incidents and problems faster. The AI can automatically detect, diagnose, respond and repair IT system abnormalities, and analyze previous events , To provide anomaly detection and recommended practices. (Full text)
Image source/Facebook, DeepMind, Google, Microsoft
Photography/Su Wenbin
AI Trends Recent News
1. Google creates an ML model that can generate song fingerprints
2. Azure Cognitive Services can now be used to detect abnormal data
3. AI education takes root downward, Gaoshi launches AI wisdom island learning platform
4. The Smart Fill function of Google Sheets is officially launched
Source: compiled by iThome, October 2020
#Trend #Weekly #Issue #Facebook #CMU #team #find #key #catalyst #turning #energy #fuel #iThome
More from Source