General machine learning systems are built from a human perspective, but researchers at the University of Washington and the Allen Institute of Artificial Intelligence tried to train AI systems with dog behavior data. The researchers collected the movement of an Eskimo dog through sensors and other devices Use the data to train the AI ​​system to achieve three goals: 1. Act like a dog and predict future actions; 2. Plan tasks like a dog; 3. Learn from dog behavior. The paper has been accepted by CVPR 2018. The significance of this work is to understand the visual data and allow the agent to take action and perform tasks. We have trained machine learning systems to recognize objects, navigate, or recognize facial expressions, but although it may be difficult, machine learning does not even reach the level of complexity that can be simulated, for example, to simulate a dog. Well, the purpose of this project is to do this-of course in a very limited way. By observing the behavior of a very well-behaved dog, this AI learned the basics of how to act like a dog. This is a collaborative study between the University of Washington and the Allen Institute of Artificial Intelligence. The paper was published at CVPR held in June this year. We studied how to directly model a visually intelligent agent. Computer vision is usually focused on solving various sub-tasks related to visual intelligence. But we deviated from this standard computer vision method; instead, we tried to directly model an agent of visual intelligence. Our model takes visual information as input and directly predicts the agent's behavior. To this end, we introduced the DECADE data set, which is a dog behavior data set collected from the dog's perspective. Using these data, we can simulate the dog's behavior and movement planning. Under a variety of measurement methods, for a given visual input, we successfully modeled the agent. In addition, compared with the representations trained by image classification tasks, the representations learned by our model can encode different information, and can also be generalized to other fields. In particular, by using this dog modeling task as a representation learning, we have obtained very good results in walkable surface estimation and scene classification tasks. Understanding visual data: imitating dogs, learning dogs Why do this research? Although there has been a lot of work on sub-tasks that simulate perception, such as identifying an object and picking it up, "understand the visual data to the point where the agent can take actions and perform tasks in the visual world." There are few studies like this. In other words, instead of simulating the behavior of the eyes, it is simulating the subject who controls the eyes. So why choose dogs? Because dogs are very complex agents, researchers say: "Their goals and motivations are often unpredictable." In other words, dogs are smart, but we don't know what they think. As a preliminary attempt in this research area, the team hopes to closely monitor the dog ’s behavior and correspond the dog ’s movements and actions to the environment it sees, to observe whether it is possible to establish an accurate prediction of these actions. system. Install a set of sensors on the body of the Eskimo Dog to collect data To achieve this, the researchers installed a basic sensor on an Eskimo dog named Kelp M. Redmon. They mounted a GoPro camera on Kelp's head, six inertial measurement units (on the legs, tail, and body) to determine the position of the object, a microphone, and an Arduino development board that tied the data together. They spent many hours recording dog activities—walking in different environments, fetching things, playing in dog parks, eating—and synchronized the dog ’s movements with the environment it saw. The result is a data set of Ego-Centric Actions in a Dog Environment (referred to as the DECADE data set). Researchers use this data set to train a new AI agent. For this agent, given some kind of sensory input—such as a scene in a room or street, or a flying ball—to predict what a dog will do in this situation. Of course, needless to say special details, even just to figure out how its body moves and where to move is already a very important task. Hessam Bagherinezhad of the University of Washington is one of the researchers. He explained: "It learned how to move joints to walk, and learned how to walk or run to avoid obstacles." "It learned to run after squirrels, followers The owner walks and chases the flying dog toy (when playing a frisbee game). These are some basic AI tasks of computer vision and robotics (such as motion planning, walkable surfaces, object detection, object tracking, person recognition), we Always try to solve it by collecting separate data for each task. " The research raises three questions: (1) imitate the behavior of the dog: predict the next behavior of the dog based on the previous behavior image of the dog; (2) plan the action like a dog; (3) learn from the behavior of the dog: for example, predict a A walkable area. These tasks can produce quite complex data: for example, a dog model must know, like a real dog, where it can walk when it needs to move from one location to another. It cannot walk on trees or cars, nor on sofas (depending on the house). Therefore, this model has learned this too. It can be deployed as a computer vision model to find out where a pet (or a footed robot) can reach in a given image. Model architecture for predicting dog behavior Model architecture for learning how dogs plan Model architecture for predicting walking areas The researchers say this is just a preliminary experiment, and although successful, the results are limited. Subsequent research may consider introducing more senses (such as smell), or seeing how a dog (or many dogs) models can be generalized to other dogs. Their conclusion is: "We hope this work will pave the way for us to better understand visual intelligence and other intelligent creatures living in our world."
The ergonomic comfortable seat minimizes the pressure on the spine caused by students sitting for a long time, and the seat is more suitable for the body, and it is more comfortable to sit for a long time. The size is upgraded, the student's range of activities is larger, the use of more comfortable oval tube, the steel tube is thicker, the weighing capacity is stronger, the service life is longer, the pen groove design prevents the pen from rolling on the table top, the side load-bearing hook, high fastness Hook design, plastic material scratch-resistant, convenient and practical. Free adjustment, lift table and chair design, to meet the needs of different heights and ages.
Kids Adjustable Desk,Lately Wooden Study Desk,Oem Kids Study Desk,Ergonomic Study Desk Igrow Technology Co.,LTD , https://www.szigrowdesks.com
