Machine learning future trends
MACHINES LEARN TO BE SAFER
How can computers that work in cars as autopilot learn the rules of the road?
Some engineers believe that it would be good if computers were learning in virtual reality, on the mistakes made there, and not on real streets.
This is a way to identify deficiencies in the operation of an autopilot, does not endanger real people. If the machine makes a mistake in simulating something, engineers can reconfigure the software by including new rules in it.
Researchers are also developing methods that will allow machines to learn new behaviors independently, it will be much faster than if engineers do it in manual coding mode.
These methods are part of a radical effort to accelerate the development of autonomous vehicles with the help of so-called machine learning.
When Google developed its first self-managed cars, nearly a decade ago, engineers created huge software in hand, carefully encoding each behavioral algorithm. But thanks to the latest improvements in processing power, autonomous automakers use sophisticated algorithms that can master tasks such as identifying pedestrians on carriageways or predicting future events on their own.
There are certain difficulties with this research, primarily they lie in the fact that algorithms process huge amounts of information that exceed the ability to process any person, and therefore it is sometimes difficult to understand the development of behavior and the causes of certain actions of programs after training. However, in the future machine learning will be of great importance for the development of autonomous vehicles.
Modern cars are still not as autonomous as we might think, but after 10 years of research, development and testing, Google cars are ready to offer a ride through the streets of Arizona. Waymo, which is a subsidiary of Google, is preparing to launch taxi services near Phoenix, and unlike other services, he will not need the help of a man, however, the machine will still remain on a short leash.
Now autonomous cars are likely to be limited to a small area, a small number of stops and a small number of pedestrians. An autonomous machine will drive at low speeds with a long wait before the maneuver, there is also a concept in which the car will try to avoid such situations and build a simpler, for autonomous implementation, route.
The leading enterprises are convinced that these cars will eventually be able to handle more complex tasks, thanks to the continued development and testing, as well as the introduction of new sensors that can provide a more detailed picture of the world around and further development of machine learning.
Waymo and many of his rivals have already embraced deep neural networks, sophisticated algorithms that can master data analysis tasks. Analyzing photos of pedestrians, for example, a neural network can learn to define a pedestrian. These algorithms also help identify road signs and markings, predict what will happen next on the road and plan the routes ahead.
The trouble is that this requires a huge amount of data collected with cameras, radar and other sensors, which requires documentation of real objects and situations. And people should create tags of this data: the identification of pedestrians, road signs and the like. Collecting, marking and describing all possible situations is impossible. It is even more difficult to find information about accidents. Simulation can help us in this.
Recently, Waymo unveiled a simulated carriageway called Carcraft. Today, the company said that this simulator provides an opportunity to test their cars on a scale that is impossible in the real world.
Toyota already uses images of simulated roads to train neural networks, and this approach has yielded promising results. In other words, modeling is close enough to the physical world.
The advantage of the simulator is that the researchers have full control over it. Do not waste time and money marking on the marking of images, with a chance of errors in the labels.
Others use a more sophisticated training method with reinforcement. This is an important line of research within many of the world's largest artificial intelligence labs, such as DeepMind (the London laboratory belongs to Google), OpenAI and others. These laboratories build algorithms that allow machines to master tasks in the virtual world through intense trial and error.
DeepMind used this method to build a machine that could play the ancient GO game better than any person. In essence, machines play thousands and thousands of games against themselves, carefully recording, every successful and unsuccessful step. And now, DeepMind and other leading world labs using similar methods in building machines that can play complex video games, such as Starcraft.
This may seem frivolous. But if machines can go into these virtual worlds, they can build their way through the physical world.
Inside the uber-autonomous car control, the researchers trained the system, play the popular Grand Theft games, with the aim of applying these methods, after all, in the real world. Learning and modeling systems in physical places is the next step.
Bridging the gap between the virtual and physical environment is not an easy task. In addition, companies must ensure that algorithms do not generate unexpected or harmful behavior during self-study on their own. This became a big concern among researchers of artificial intelligence.
For this and other reasons, companies such as Toyota and Waymo do not build autonomous vehicles solely around machine learning. They still use hand-coded software, these more traditional ways will help provide certain behaviors. Therefore, the machines of Waymo will not learn to stop at stop lamps, it will be spelled out in manual mode as a rigid rule.
However, the industry is moving towards machine learning, and this is the best way to train stand-alone machines to perform complex tasks, the performance of which requires a much deeper understanding of the world around them.
for more detail : www.digitaltechnologyreview.com
How can computers that work in cars as autopilot learn the rules of the road?
Some engineers believe that it would be good if computers were learning in virtual reality, on the mistakes made there, and not on real streets.
This is a way to identify deficiencies in the operation of an autopilot, does not endanger real people. If the machine makes a mistake in simulating something, engineers can reconfigure the software by including new rules in it.
Researchers are also developing methods that will allow machines to learn new behaviors independently, it will be much faster than if engineers do it in manual coding mode.
These methods are part of a radical effort to accelerate the development of autonomous vehicles with the help of so-called machine learning.
When Google developed its first self-managed cars, nearly a decade ago, engineers created huge software in hand, carefully encoding each behavioral algorithm. But thanks to the latest improvements in processing power, autonomous automakers use sophisticated algorithms that can master tasks such as identifying pedestrians on carriageways or predicting future events on their own.
There are certain difficulties with this research, primarily they lie in the fact that algorithms process huge amounts of information that exceed the ability to process any person, and therefore it is sometimes difficult to understand the development of behavior and the causes of certain actions of programs after training. However, in the future machine learning will be of great importance for the development of autonomous vehicles.
Modern cars are still not as autonomous as we might think, but after 10 years of research, development and testing, Google cars are ready to offer a ride through the streets of Arizona. Waymo, which is a subsidiary of Google, is preparing to launch taxi services near Phoenix, and unlike other services, he will not need the help of a man, however, the machine will still remain on a short leash.
Now autonomous cars are likely to be limited to a small area, a small number of stops and a small number of pedestrians. An autonomous machine will drive at low speeds with a long wait before the maneuver, there is also a concept in which the car will try to avoid such situations and build a simpler, for autonomous implementation, route.
The leading enterprises are convinced that these cars will eventually be able to handle more complex tasks, thanks to the continued development and testing, as well as the introduction of new sensors that can provide a more detailed picture of the world around and further development of machine learning.
Waymo and many of his rivals have already embraced deep neural networks, sophisticated algorithms that can master data analysis tasks. Analyzing photos of pedestrians, for example, a neural network can learn to define a pedestrian. These algorithms also help identify road signs and markings, predict what will happen next on the road and plan the routes ahead.
The trouble is that this requires a huge amount of data collected with cameras, radar and other sensors, which requires documentation of real objects and situations. And people should create tags of this data: the identification of pedestrians, road signs and the like. Collecting, marking and describing all possible situations is impossible. It is even more difficult to find information about accidents. Simulation can help us in this.
Recently, Waymo unveiled a simulated carriageway called Carcraft. Today, the company said that this simulator provides an opportunity to test their cars on a scale that is impossible in the real world.
Toyota already uses images of simulated roads to train neural networks, and this approach has yielded promising results. In other words, modeling is close enough to the physical world.
The advantage of the simulator is that the researchers have full control over it. Do not waste time and money marking on the marking of images, with a chance of errors in the labels.
Others use a more sophisticated training method with reinforcement. This is an important line of research within many of the world's largest artificial intelligence labs, such as DeepMind (the London laboratory belongs to Google), OpenAI and others. These laboratories build algorithms that allow machines to master tasks in the virtual world through intense trial and error.
DeepMind used this method to build a machine that could play the ancient GO game better than any person. In essence, machines play thousands and thousands of games against themselves, carefully recording, every successful and unsuccessful step. And now, DeepMind and other leading world labs using similar methods in building machines that can play complex video games, such as Starcraft.
This may seem frivolous. But if machines can go into these virtual worlds, they can build their way through the physical world.
Inside the uber-autonomous car control, the researchers trained the system, play the popular Grand Theft games, with the aim of applying these methods, after all, in the real world. Learning and modeling systems in physical places is the next step.
Bridging the gap between the virtual and physical environment is not an easy task. In addition, companies must ensure that algorithms do not generate unexpected or harmful behavior during self-study on their own. This became a big concern among researchers of artificial intelligence.
For this and other reasons, companies such as Toyota and Waymo do not build autonomous vehicles solely around machine learning. They still use hand-coded software, these more traditional ways will help provide certain behaviors. Therefore, the machines of Waymo will not learn to stop at stop lamps, it will be spelled out in manual mode as a rigid rule.
However, the industry is moving towards machine learning, and this is the best way to train stand-alone machines to perform complex tasks, the performance of which requires a much deeper understanding of the world around them.
for more detail : www.digitaltechnologyreview.com
Comments
Post a Comment