These Neural Networks Know What They’re Doing

Certain types of artificial intelligence agents can learn the basics of causes and consequences of navigation tasks during training.

neural network You can learn to solve all kinds of problems, from identifying cats in pictures to driving self-driving cars. However, whether these powerful pattern recognition algorithms actually understand the task they are performing remains an open question.

For example, a neural network whose mission is to keep self-driving cars in the lane learns to do so by looking at the bushes beside the road, rather than detecting the lane and focusing on the horizon of the road. maybe.

Researcher MIT This shows that certain types of neural networks can learn the true causal structure of navigation tasks that are trained to perform. These networks can understand tasks directly from visual data, so they should be more effective than other neural networks when navigating in complex environments such as densely populated areas or rapidly changing weather conditions. ..

In the future, this task could improve the reliability and reliability of machine learning agents performing high-stakes tasks, such as driving self-driving cars on congested highways.

Deep learning neural network navigation

MIT researchers have demonstrated that a special class of deep learning neural networks can learn the true causal structure of navigation tasks during training.Credit: Stock image

“These machine learning systems can perform inferences in a causal way, so we can know and point out how they work and make decisions. This is essential for safety-critical applications. “” Said Ramin Hasani, co-author of the Computer Science and Artificial Intelligence Laboratory (CSAIL) postdoc.

Co-authors include Charles Vorbach, a graduate student in electrical engineering and computer science and co-author. CSAIL Doctoral Student Alexander Amini; Graduate Student of the Austrian Institute of Science and Technology Mathias Lechner; Senior Author Daniela Rus, Professor of Electrical Engineering and Computer Science at Andrew and Erna Bitervi, Director of CSAIL. The study will be presented at the 2021 Conference on Neural Information Processing Systems (NeurIPS) in December.

Attention results

Neural networks are a method of machine learning in which a computer learns to complete a task through trial and error by analyzing many training examples.When “Liquid” neural network Modify the underlying equation to continuously adapt to the new input.

A new study is how Hasani et al.’S brain-inspired type of deep learning system Neural circuit policy Constructed by a liquid neural network cell (NCP), a network of only 19 control neurons can autonomously control an autonomous vehicle.

Researchers observe that NCPs performing lane keeping tasks continue to pay attention to road horizons and boundaries when making driving decisions, much like humans are driving a car. bottom. The other neural networks they studied were not always focused on the road.

“It was a great observation, but we didn’t quantify it, so we wanted to find a mathematical principle of why and how these networks could capture the true causality of the data,” he said. Says.

They found that when the NCP was trained to complete a task, the network interacted with the environment and learned to explain the intervention. In essence, the network recognizes whether its output has been altered by a particular intervention and correlates cause and effect.

During training, the network runs forward to produce output and then runs backwards to fix the error. Researchers have observed that NCP correlates cause and effect between forward and backward modes. This allows the network to focus very much on the true causal structure of the task.

Hasani and his colleagues did not have to impose additional constraints on the system or perform any special settings on the NCP to learn this causal relationship.

“Causality is especially important for characterizing applications where safety is important, such as flight,” Rus says. “Our work shows the causal link of neural circuit policies for in-flight decision making, including flight and formation flight in densely obstructed environments such as forests.”

Changes in the weathered environment

They tested NCP through a series of simulations of autonomous drones performing navigation tasks. Each drone navigated using input from one camera.

The drone was responsible for moving to the target object, tracking the moving target, or tracking a series of markers in different environments such as Redwood Forest and neighborhoods. They also traveled under a variety of weather conditions, including sunny weather, heavy rain, and fog.

Researchers have found that on sunny days, NCPs work on simple tasks like any other network, but outperform all networks on more difficult tasks, such as chasing moving objects in a storm. I did.

“We found that NCP is the only network that pays attention to objects of interest in different environments when testing different locations, lighting, and environmental conditions when completing navigation tasks. It’s the only system that can easily do this and actually learn what the system is trying to learn, “he says.

Their results show that NCP allows autonomous drones to navigate successfully even in changing conditions, such as sunny landscapes with sudden fog.

“Learning what the system really does can work well in new scenarios and environmental conditions that we have never experienced before. This is a major challenge for today’s non-causal machine learning systems. I believe these results are very exciting because they show how causality emerges from the choice of neural network, “he says.

In the future, researchers want to find ways to use NCPs to build larger systems. Bringing together thousands or millions of networks allows you to tackle even more complex tasks.

See also: “Causal Navigation with Continuous Time Neural Networks” by Charles Vorbach, Ramin Hasani, Alexander Amini, Mathias Lechner, Daniela Rus, June 15, 2021 Computer Science> Machine Learning..
arXiv: 2106.08314

This study was supported by the US Air Force Research Laboratory, the US Air Force Artificial Intelligence Accelerator, and Boeing.

These Neural Networks Know What They’re Doing Source link These Neural Networks Know What They’re Doing

Related Articles

Back to top button