Some type of artificial intelligence agent can learn the cause and effect basis of a navigation task during training.
Neural networks can be taught to solve all kinds of problems, from identifying cats in photos to driving an autonomous car. But whether these powerful pattern recognition algorithms actually understand the tasks they perform remains an open question.
For example, a neural network tasked with keeping an autonomous car in its lane might learn how to do this by observing the bushes by the side of the road, rather than learning to sense lanes and focus on the horizon of the road. .
Researchers from MIT have now shown that a certain type of neural network is capable of learning the true cause and effect structure of the navigation task for which it is trained. Since these networks can understand the task directly from visual data, they should be more efficient than other neural networks when navigating in a complex environment, such as a location with dense trees or rapidly changing weather conditions.
In the future, this work could improve the reliability and reliability of machine learning agents who perform high-stakes tasks, such as driving an autonomous vehicle on a busy highway.
“Because these machine learning systems are able to reason causally, we can know and indicate how they work and make decisions. This is essential for security-critical applications, ”explains co-lead author Ramin Hasani, post-doctoral fellow at the Computer Science and Artificial Intelligence Laboratory (CSAIL).
Co-authors include a graduate student in electrical engineering and computer science and lead co-author Charles Vorbach; CSAIL doctoral student Alexander Amini; Mathias Lechner, graduate student of the Austrian Institute of Science and Technology; and lead author Daniela Rus, Andrew and Erna Viterbi Professor of Electrical Engineering and Computer Science and Director of CSAIL. The research will be presented at the 2021 Neural Information Processing Systems (NeurIPS) conference in December.
An eye-catching result
Neural networks are a method of machine learning in which the computer learns to perform a task through trial and error by analyzing many training examples. And “liquid” neural networks change their underlying equations to constantly adapt to new inputs.
The new research builds on earlier work in which Hasani and others have shown how a type of brain-inspired deep learning system called Neural Circuit Policy (NCP), built by liquid neural network cells , is capable of autonomously controlling an autonomous vehicle, with a network of only 19 control neurons.
The researchers observed that NCPs performing a lane keeping task kept their attention on the horizon and road boundaries when making a driving decision, much like a human would (or should). while driving a car. The other neural networks they studied did not always focus on the road.
“It was an interesting observation, but we didn’t quantify it. So we wanted to find the mathematical principles explaining why and how these networks are able to capture the true causality of the data, ”he says.
They found that when a NCP is trained to perform a task, the network learns to interact with the environment and to report on interventions. Essentially, the network recognizes if its output is altered by some intervention, and then ties the cause and effect together.
During training, the network is run forward to generate output, and then backward to correct errors. The researchers observed that PCNs establish a cause-and-effect link in forward and reverse mode, which allows the network to pay very focused attention to the true causal structure of a task.
Hasani and his colleagues did not need to impose additional constraints on the system or perform any special configuration for the NCP to learn this causation.
“Causality is particularly important to characterize for safety critical applications such as theft,” says Rus. “Our work demonstrates the causal properties of neural circuit policies for in-flight decision making, including flight in environments with dense obstacles such as forests and formation flight.”
Alteration of environmental changes
They tested the PCNs through a series of simulations in which autonomous drones performed navigation tasks. Each drone used inputs from a single camera to navigate.
The drones were tasked with getting to a target object, chasing a moving target, or following a series of markers in a variety of environments, including a redwood forest and a neighborhood. They also traveled in different weather conditions, such as clear skies, heavy rain and fog.
The researchers found that NCPs performed as well as other networks on simpler tasks in good weather, but outperformed them all on more difficult tasks, such as chasing a moving object through a rainstorm. .
“We have observed that PCNs are the only network that pays attention to the object of interest in different environments while performing the task of navigation, wherever you test it, and in different lighting or environmental conditions. It’s the only system that can do this casually and actually learn the behavior we want the system to learn, ”he says.
Their results show that using NCP could also allow autonomous drones to successfully navigate environments with changing conditions, such as a sunny landscape that suddenly becomes foggy.
“Once the system learns what it is actually supposed to do, it can perform well in new scenarios and environmental conditions that it has never experienced. This is a big challenge with today’s non-causal machine learning systems. We think these results are very exciting, because they show how causality can emerge from the choice of a neural network, ”he says.
In the future, researchers want to explore the use of PCNs to build larger systems. Setting up thousands or millions of networks could allow them to tackle even more complex tasks.
Reference: “Causal Navigation by Continuous-time Neural Networks” by Charles Vorbach, Ramin Hasani, Alexander Amini, Mathias Lechner and Daniela Rus, June 15, 2021, Computer Science> Machine Learning.
This research was supported by the United States Air Force Research Laboratory, the US Air Force Artificial Intelligence Accelerator, and the Boeing Company.