The research in machine learning և AI, which is now a key technology in virtually every field և company, is too extensive for anyone to read it all. This column is Perceptron (formerly: Deep science) aims to collect the most recent discoveries և documents, particularly on artificial intelligence, but not limited to և explain why they are important.
This week AI, researchers was found a method that can allow opponents to track the movements of remote-controlled robots even when the robots’ communication is encrypted from end to end. The co-authors, originally from the University of Strathclyde in Glasgow, say their research shows that adopting best practices in cybersecurity is not enough to stop attacks on autonomous systems.
Remote control or remote control promises to enable operators to control one or more robots remotely in different environments. Startups, including: Pollen Robotics:, Beamև: Tortoise demonstrated the usefulness of remote-controlled robots in grocery stores, hospitals and offices. Other companies are developing remote-controlled robots for tasks such as destroying bombs or exploring areas of strong radiation.
But new research shows that remote control, even when supposedly “safe”, is risky because of its susceptibility to control. The co-authors of Strathclyde describe in a piece of paper how the remote-controlled robot performs information through a neural network. After collecting samples TLS:Analyzing the protected traffic between the robot and the controller, they found that the neural network could detect movements in about 60% of the time and restore “storage workflows” (such as picking packets) with “high accuracy.”
Anxiety about a less direct egg is a novelty study By Google և Researchers at the University of Michigan, who studied people-to-people relationships with AI systems with weak legislation և in countries with “popular optimism” for AI. The study was conducted by users of “financially depressed” instant credit platforms in India, which target borrowers with a loan designed with artificial intelligence risk modeling. According to the co-authors, users felt indebted to the “benefits” of instant loans – to accept strict terms, to overuse sensitive data – to pay high fees.
The researchers argue that the findings show the need for greater “algorithmic accountability”, especially when it comes to financial services AI. “We argue that accountability is shaped by a platform-user power relationship; we urge policymakers to be careful about adopting a purely technical approach to promoting algorithmic accountability,” they wrote. “Instead, we call for local interventions that strengthen the user agency, enable meaningful transparency, restructure the designer-user relationship, and encourage interns to critically reflect on broader accountability.”
Less in misery researchA team of Dutch scientists from TU Dortmund University, Rhine-Waal University և LIACS Universiteit Leiden has developed an algorithm that they believe can “solve” the Rocket League game. Aiming to find a less computationally intensive way to create game AI, the team used what they call a “sim-to-sim” transfer technique that trained the AI system to perform in-game tasks such as goalkeeper և hitting. A simplified, simplified version of Rocket League. (Rocket League is basically like indoor soccer, except for cars instead of players from three-man teams).
It was not perfect, but the researchers’ Rocket League-playing system was able to save almost all the shots fired by the goalkeeper. During the attack, the system successfully hit 75% of the strokes, which is a respectable record.
Human movement simulators are also evolving rapidly. Meta’s work on tracking and modeling on human limbs is obviously used in its AR և VR products, but it can also be more widely used in robotics-embodied AI. The study, which came out this week, got the tip of the iceberg none other than Mark Zuckerberg.
MyoSuite: mimics muscles and skeletons in 3D as they interact with objects և. This is for agents to learn how to properly handle and manage items without crushing or dropping them, as well as providing realistic catches and interactions in the virtual world. It is supposed to work on certain tasks thousands of times faster, which allows the simulated learning process to be done much faster. “We are developing these open source models so that researchers can use them to further the field,” says Zook. And they did.
Many of these simulations are based on agents or objects, but this project from MIT looks at the general system of independent agents, modeling self-driving cars. The idea is that if you have a lot of cars on the road, you can force them to work together not only to avoid collisions, but also to avoid idle stops at unnecessary lights.
As you can see in the animation above, some autonomous vehicles that communicate via v2v protocols can generally prevent all but the front cars from stopping, gradually slowing down one after the other, but not so much that they actually stop. . This kind of hypermiliary behavior may not seem like it saves a lot of gas or battery, but when you measure it up to thousands or millions of cars, it makes a difference; it can also be more comfortable. Good luck getting everyone to the intersection, which is at a perfect distance, though.
Switzerland looks good for a long time using 3D scanning technology. The country makes a huge map using drones equipped with a lid and other tools, but there is one thing: the movement of the drone (deliberate և accidental) makes a mistake in the point map, which must be corrected manually. Isn’t it a problem if you are only looking at one building but a whole country?
Fortunately, the team outside EPFL integrates the ML model directly into the Leader capture stack, which can determine when an object has been scanned multiple times from different angles, and use that information to arrange the point map into a single network. This news article not particularly enlightening, but The accompanying paper is more detailed. A copy of the resulting map can be seen in the video above.
Finally, the team from the University of Zurich received some unexpected but very good AI news designed an algorithm to track animal behavior so zoologists will not have to clean up weeks of footage to find two copies of amateur dance. It is a cooperation with the Zurich Zoo, which makes sense if we take into account the following: “Our method can detect even subtle or infrequent changes in the behavior of the animals being studied, such as signs of stress, anxiety or restlessness,” said laboratory director Mehmet Fatih Yanik.
Thus, the tool can be used for both captive learning and tracking behavior, as well as for the welfare of captive animals in zoos, as well as for other study animals. They could use fewer animals, get more information in less time, have less work done by classmates, and watch video files until late at night. It sounds like a win-win-win situation for me.
Love the illustration!