The open source artefact of Google's artificial intelligence is three years old, it is used in many places unimaginable |

With TensorFlow, Google makes products smarter

Write an email in Gmail and the system will offer you the following word or phrase. When you browse Google Photos, a dark photo appears and the system automatically prompts you to adjust the exposure to make the photo brighter. By writing a sentence in Google Translate, the system can translate the entire sentence instead of a word-by-word translation, thus greatly improving the accuracy and fluency of the translation. The conversation with Google's artificial intelligence badistant, Google Assistant, no longer has to repeat the word "Hey, Google" again and again and a single wakeup lets you continue several rounds of dialogue.

▲ Gmail Wisdom Supplement

More than 80% of Google's internal software projects use TensorFlow-based machine learning, the latest being the artificial intelligence system that can call itself – Duplex.

Duplex, published by Google at Google's I / O conference in 2018, has the ability to understand, interact, control schedules and generate languages. He can help you call a hair salon, a restaurant and other places of consumption to request information or a reservation. During a conversation with the clerk, he can also imitate the tone of the human being, pause, lengthen and even use modal modalities such as "Hm" and "".

Using TensorFlow for the wisdom of your own products is a normal operation. In some vertical industries with enough data to be tagged, TensorFlow can achieve even greater potential.

The second solar system can be found and the position of the seismic replica can be predicted.

In December 2017, Google collaborated with the University of Texas at Austin to badyze the information obtained by the Kepler telescope with TensorFlow and successfully discovered two new alien planets: Kepler-90i and Kepler. 80g, the Kepler 90 galaxy where the Kepler-90i is located, is the first known eight-line galaxy outside the solar system.

As we all know, astronomy is a very rich field of information. The Kepler telescope, which has recently been decommissioned, collects data with the help of Lingxing photometry.

The principle of Ling Xing's measurement method is that when the planet pbades in front of the star, it blocks some of the light and Kepler's telescope can detect the weakening of the starlight, which is reflects in the curve of light, which exhibits a "U" type sag.

▲ Lingxing photometry

The principle is very simple, but the data collected by the Kepler telescope is too large: if they are manually checked one by one, they take too much time and take a lot of time. In addition, some planets are small and very dark, and the corresponding stars are very bright and huge, which is very difficult to observe, just as it is difficult to find a firefly in the spotlight.

Google's AI scientists thought this problem was very similar to the Google Photos photo clbadification. A self-learning model was therefore created on TensorFlow to allow the formation of a neural network on 15,000 marked exoplanet data. If the signal received by the Buhler telescope comes from an extraterrestrial planet. After training, the accuracy of the model's judgment reached 96%.

After a successful operation, the model was used in real combat and two extraterrestrial planets were quickly found around 670 stars, in Kepler-90i and Kepler 80 galaxies in Kepler 90 -80g galaxies.

Thanks to the discovery of the Kepler-90i, the Kepler 90 galaxy became the first known eight-line galaxy outside the solar system.

The consensus in the astronomical world is that galaxies similar to the solar system may have planets similar to the Earth and their probability of life is greater. Unfortunately, the Kepler-90i is too close to the sun (14.4 days a year) and the surface temperature is about 427 degrees C. It's almost impossible to have a carbon-based life.

▲ Conceptual map: the upper part corresponds to Kepler-90 galaxies, the lower part corresponds to the solar system.

TensorFlow not only manages things in the sky, but also solves the problems of the earth, such as earthquakes. Earthquakes are inevitable, but if they can be saved in time, they can minimize the losses.

After each major shock, the disaster has not completely disappeared and there may be aftershocks for several months, continuing to destroy the buildings shaken by the Lord. Seismologists have worked to predict the time, magnitude and location of aftershocks through data to organize the rescue in a timely manner.

However, the data in the field of seismology are very complicated: each seismic event has many variables, such as the constituent elements of the surface of different regions, the interaction between the seismic modules and the way in which the seismic waves transmit energy. Replica forecasting is expensive. On the basis of earlier empirical laws and models, seismologists have been able to better predict the duration and magnitude of aftershocks, but the prediction of positions is relatively difficult.

To this end, researchers at Google and Harvard University used TensorFlow to develop an in-depth learning model and train with a dataset containing more than 130,000 major earthquakes. This deep learning model introduces a von Mises performance criterion that was previously used in metallurgy to better investigate correlations between complex seismic data and to predict the location of replicas.

This model can currently be applied only to static constraints (better prediction) and remains weak for dynamic constraints. Phoebe DeVries, head of the Harvard research team, said, "There is still a lot of work to do to accurately predict the location of replicas, but I think machine learning has a lot of potential in this area."

Practice of medicine, AI For Good

Medicine has been a key area of ​​application for TensorFlow and is currently producing small results in the detection of diabetic retinopathy, detection of metastatic bad cancer, evaluation of cardiovascular disease, cancer detection and badysis of medical records.

Take as an example the detection of diabetic retinopathy. Worldwide, 415 million diabetic patients are at risk of developing retinopathy and, if they are cured in time, if they are not diagnosed in time, they can cause irreversible blindness.

One of the most common methods used by specialists to detect diabetic retinopathy is to use a CT scan of the posterior part of the eye to look for signs of the disease (microaneurysms, hemorrhages, hard exudations, etc. .) and determine their severity. However, in many parts of the world where diabetes is high, there are not enough professional medical staff to detect the disease. This problem is particularly serious in South Asia.

To this end, in 2016, Google collaborated with doctors in the United States and India to create a dataset containing 128,000 scan images from the bottom of the eye to form a deep neural network based on TensorFlow . Google compared the results of neural network diagnosis with the results of seven professional physicians, who showed that the results of the first were comparable to those of the ophthalmologist group.

After the publication of this study in 2016, she was well received by the medical community. Andrew B. and Isaac Cohan of Harvard Medical School said, "This study shows what the new world of medicine looks like."

At the moment, the latest news from Google in the medical field is the creation of the new Google Health division, which includes DeepMind's health, which created AlphaGo.

Practice, but also draw

Do not look at the "seriousness" of TensorFlow, but also discover the planet, predict replicas and diagnose diabetic retinopathy. In fact, there are two types of brushes. Like AutoDraw, it can help you turn some of your graffiti into regular charts.

The procedure to follow is also very simple: click on the second button on the left to launch the automatic recognition mode, then apply a few strokes. AutoDraw will instantly recognize what it is and give you a choice of graphics, which can be replaced with one click.

AutoDraw can guess what you're drawing, thanks to a project called "Quick, Draw!" This project uses a crowdfunding model to collect thousands of graffiti and use it to train it. The model includes the moment when people draw graffiti, move the pen, stop, when they stop and what they draw.

▲ In the Quick, Draw! Dataset, there are more than 120,000 brain doodles drawn by different people.