Future Prediction Systems


( c )


The best neural networks that beat a person in poker, Go, chess and DotA have one thing in common - they can predict the near future.


The ability of machines to predict behavior can far exceed human capabilities. In the space of various probabilities, algorithms are better than a person who is affected by emotions.


What can neural networks predict? We are faced with an endless field of opportunities: the exchange, crime, weather, health, transport - everywhere the ability to calculate a few steps forward will be useful. Already today, some algorithms are superior to human experts. The dawn of tomorrow's neurodnya will not leave any trace of the "fog of the unknown."


Researchers from DeepMind published a scientific paper in which they presented a new method of reinforcing a neural network. It turned out that if in the process of self-learning the neural network begins to “ fantasize ” about various options for the future, then it learns much faster. The “fantasy” of the neural network is that, according to the last three known frames, the neural network should predict the reward that it will receive in the fourth unknown time interval. AI uses its memory and applies new strategies as if in its imagination.


The more efficient systems become, the better they make forecasts. Now we can not only predict the weather (in the short term). We can even “see” the future of macroeconomic situations in various areas of the city, measuring the consumption of water, electricity, traffic (how many passengers are in public transport and how many in our cars), increase / decrease in resource consumption.


It is already difficult to imagine a sphere in which we could do without predictions. And is it worth it to abandon them if the algorithms make it possible to choose the right behavior strategy?


Road behavior



Researchers at the Massachusetts Institute of Technology have built a system that can predict a huge number of real-world events. At first, the program was trained on a choice of 2 million online videos. The program analyzed each video, classifying all objects and actions in the plots.


Then the neural networks showed a static image. The program, in turn, generated 1.5-second video clips that demonstrated a vision for the near future.


Obviously, such a solution can be used not only to create GIFs. Algorithms, in principle, allow you to "look" into the future of complex systems, which will find application in autonomous cars that analyze an ever-changing situation on the road.


The computer will be able to understand that it sees something unusual - for example, an animal ran out onto the road. Even if the car never got into this situation before, it will "understand" that something strange is happening - you should either stop or transfer control to the driver.


Human health



( c )


Scientists at Stanford University have developed an artificial intelligence system that can predict the likelihood of death of a seriously ill patient within a year with an accuracy of 90%.


Researchers analyzed records of 160,000 patients to collect data on past diagnoses, prescribed procedures, and predictions made by doctors.


After processing the dataset, an algorithm was compiled for deep learning of the neural network. The grid then predicted all-cause mortality for a period of 3 to 12 months for 40,000 patients.


A year later, the researchers summarized: in 90% of cases, the neural network correctly predicted the condition of the patient (regardless of whether he was waiting for his death or recovery). This indicator significantly exceeds the capabilities of even a group of medical experts.


The Teraflu brand has developed a system that predicts the likelihood of catching a cold in several countries, including Russia. Every day, the system analyzes posts in social networks, queries in search engines, data from the “ Research Institute of Influenza ”, as well as data on the demand in pharmacies for funds specific to the fight against cold symptoms. The result is a graph of the "catarrhal danger" in a particular region with a forecast for several days. However, such platforms find more valuable application: in the Virtual Singapore system, you can now in real time, view and analyze the country's life and predict, for example, the spread of dangerous infections or the reaction of large masses of people to an explosion in a shopping center.



Microsoft and Adaptive Biotechnologies plan to create a system that, based on a blood test, will be able to detect diseases in the early stages. By analyzing the genetic code in trillions of T-lymphocyte receptors, the system will identify the diseases that the body has encountered at an asymptomatic stage. The assumption is that the test will be able to detect a wide range of diseases at a time, including diseases that are usually diagnosed in very late stages.


A research group from the Institute of Molecular Biology of the Russian Academy of Sciences, the Russian Gerontological Scientific and Clinical Center, Moscow Institute of Physics and Technology and other research centers presented a method for predicting a person’s biological age (which differs from the passport age) based on ultrasound data of the human carotid artery and tonometry. Using machine learning , a complex formula was obtained that is able to predict the age of healthy people with an accuracy of 6.9 years for men and 5.9 years for women, which is a very high indicator compared to other known methods.


Danish scientists have developed the Corti Signal neural network, which tracks audio messages to diagnose a heart attack. First of all, the system should help people who called the ambulance. The operator is not always able to detect a heart attack in a person on the other end of the wire (cope in 73% of cases), but the neural network solves this problem with an accuracy of 95%! The AI ​​not only listens to the conversation, but also collects non-verbal signals, such as breathing patterns.


Apparently, in the future, systems based on neural networks (and other methods) will make it possible to predict diseases much earlier - in some cases, decades before the onset of the disease itself.


Smart things know what will happen to them



Imagine a building that, even before the accident, can say that, for example, heating will soon fail. Some companies use machine learning to do just that. This procedure is called predictive maintenance.


CGnal, based in Milan, Italy, recently analyzed annual data from heating and ventilation systems in an Italian hospital. From the sensors, data were obtained on temperature, humidity, and electricity use. The algorithm was trained on the sample for six months, then the researchers checked it according to the data from the second half of the year. The system predicted 76 out of 124 real faults, including 41 out of 44, where the temperature of the instrument rose above acceptable levels.


Other companies also use a similar approach to data. Finnish startup Leanheat places a wireless temperature, humidity and pressure sensor for remote heating control and device health monitoring. Instead of regulating the heating simply by the outside temperature, Leanheat models take into account weather changes: the temperature dropped to zero from 10 degrees or rose from -10.


In the US, Augury developed Shazam for Machines by installing acoustic sensors in machines to listen for audible changes and identify potential imminent failures. However, the gadget can work with different devices: customers can connect the sensor to commercial refrigerators or industrial heaters. The Augury gadget records vibrations and ultrasounds, uploads them to the cloud service, where the data is analyzed to forecast the performance of the controlled machine.


Audio and data are analyzed and stored so that the sound of one client device can be compared with the sound of all the others. The idea is that Augury does not need to configure software for each type of device. Instead, you can simply install the sensors and listen to the device to create an idea of ​​how it sounds when it functions normally. Over time, the sound database will let you know which specific sounds precede specific types of malfunctions.


Weather forecast



Weather forecasting remains a challenge for science. We already got the hang of using convolutional neural networks for this, but progress does not stand still. In the list of Top-500 most powerful computing systems in the world, as of November 2016, 23 supercomputers were engaged in weather forecasting.


ClimaCell uses an approach that is not related to neural networks and super-complex algorithms: wireless communication networks act as weather prediction sensors - all this is done within the framework of the nowcasting concept, in which an ultra-short-term forecast of weather phenomena is made within 0-6 hours from the observation period .


ClimaCell combines several levels of data from wireless networks, satellites, weather radars and other sensors to create high-definition maps. Using data from approximately 5,000 stations operated by several telecommunications companies, the company creates very accurate and reliable weather maps.


Dangerous Algorithms



Not that compass , but close in meaning


Various crime prediction systems have been tested in the USA for several years. One of the first systems of this type - COMPAS - was created in 1998. COMPAS analyzes 137 biographical parameters of a convicted person, including the severity of previous crimes, level of education and income, marital status and addictions. The program also takes into account the results of psychological tests, including temperament, risk appetite, the degree of narcissism and a tendency to guilt. Based on these data, COMPAS predicts the likelihood of a criminal relapse in the next two years.


However, at Dartmouth College, they conducted a thorough COMPAS study and concluded that the algorithm is actually no more accurate than any average person. The program was able to identify repeat offenders in 65% of cases. People without special education and experience in sentencing coped with this task in 67% of cases, knowing only the age, gender and history of the crimes of the accused. Moreover, it turned out that COMPAS accuracy can be improved if only two parameters are left in it: a person’s age and information about previous convictions.


Algorithms can make decisions and make forecasts much more efficiently than humans. People take into account non-essential factors and ignore the really important ones, give in to emotions, and also allow themselves to make decisions in accordance with their internal “instinct”, intuition, or without any logic at all.


However, this does not mean that we should completely trust the machines, because they also do not have 100% accuracy.