Neural networks can learn to solve a wide range of problems, from recognising cats in photos to steering a self-driving car. However, it is unclear whether these powerful pattern-recognition algorithms truly understand the tasks they are performing.For example, instead of learning to detect lanes and focus on the road’s horizon, a neural network tasked with keeping a self-driving car in its lane might learn to do so by watching the bushes at the side of the road.MIT researchers have demonstrated that a specific type of neural network can learn the true cause-and-effect structure of the navigation task it is being trained to perform. Because these networks can understand the task directly from visual data, they should outperform other neural networks when navigating in a complex environment, such as one with dense trees or rapidly changing weather conditions.
This research could improve the dependability and trustworthiness of machine learning agents performing high-stakes tasks, such as driving an autonomous vehicle on a busy highway, in the future.”Because these machine-learning systems can perform causal reasoning, we can understand and point out how they function and make decisions.” This is critical for safety-critical applications,” says Ramin Hasani, a postdoctoral researcher at the Computer Science and Artificial Intelligence Laboratory (CSAIL).Charles Vorbach, a graduate student in electrical engineering and computer science and co-lead author, CSAIL PhD student Alexander Amini, Institute of Science and Technology Austria graduate student Mathias Lechner, and senior author Daniela Rus, the Andrew and Erna Viterbi Professor of Electrical Engineering and Computer Science and director of CSAIL, are among the co-authors. The findings will be presented in December at the 2021 Conference on Neural Information Processing Systems (NeurIPS).
An eye-catching outcomeNeural networks are a machine learning method in which the computer learns to complete a task through trial and error by analysing a large number of training examples. In addition, “liquid” neural networks modify their underlying equations in order to continuously adapt to new inputs.
The new study builds on previous work by Hasani and colleagues, who demonstrated how a brain-inspired type of deep learning system known as a Neural Circuit Policy (NCP) built from liquid neural network cells can autonomously control a self-driving vehicle with a network of only 19 control neurons.When performing a lane-keeping task, the NCPs kept their attention on the road’s horizon and borders when making a driving decision, just as a human would (or should) while driving a car. Other neural networks they investigated were not always focused on the road.
“That was an interesting observation, but we did not quantify it.” “As a result, we wanted to discover the mathematical principles underlying why and how these networks are capable of capturing the true causation of the data,” he says.They discovered that when an NCP is trained to perform a task, the network learns to interact with its surroundings and account for interventions. In essence, the network detects when its output is altered by a specific intervention and then connects the cause and effect.The network is trained by running it forward to generate an output and then backward to correct for errors.
The researchers discovered that NCPs relate cause-and-effect during forward and backward modes, allowing the network to focus its attention on the true causal structure of a task.For the NCP to learn this causality, Hasani and his colleagues did not need to impose any additional constraints on the system or perform any special setup.”Causality is particularly important to characterise in safety-critical applications like flight,” says Rus. “Our research demonstrates the causality properties of Neural Circuit Policies for decision-making in flight, including flying in densely obstructed environments such as forests and flying in formation.”coping with environmental changesThey put NCPs to the test by running a series of simulations in which autonomous drones performed navigation tasks. To navigate, each drone relied on input from a single camera.The drones had to travel to a target object, chase a moving target, or follow a series of markers in a variety of environments, including a redwood forest and a neighbourhood.
They also travelled in a variety of weather conditions, including clear skies, heavy rain, and fog.The researchers discovered that the NCPs performed similarly to the other networks in good weather, but outperformed them all in more difficult tasks, such as chasing a moving object through a rainstorm.
“We discovered that NCPs are the only networks that pay attention to the object of interest while completing the navigation task, regardless of where you test it or the lighting or environmental conditions.”
“This is the only system that can do this casually while also learning the behaviour we want the system to learn,” he says.Their findings indicate that the use of NCPs could also allow autonomous drones to successfully navigate in environments with changing conditions, such as a sunny landscape that suddenly becomes foggy.”Once the system understands what it is supposed to do, it can perform well in novel scenarios and environmental conditions that it has never encountered before.” This is a significant challenge for current non-causal machine learning systems. “We think these findings are very exciting because they show how causality can emerge from the selection of a neural network,” he says.In the future, the researchers hope to investigate the use of NCPs in the construction of larger systems.
By connecting thousands or millions of networks, they may be able to tackle even more difficult tasks. The United States Air Force Research Laboratory, the United States Air Force Artificial Intelligence Accelerator, and the Boeing Company all contributed to this study.
_______
Network | Don’t forget to follow us on Twitter @njtimesofficial. To get the latest updates
