A certain type of artificial intelligence, called neural networks, have been found to be able to learn by cause and effect analysis. “For example, 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, rather than learning to detect the lanes and focus on the road’s horizon,” says an article written by MIT.
This certain type of neural network is beginning to learn the cause and effect structure of a navigation task. Because they learn from visual data, researchers project these neural networks to be the best for the job of navigating a complex environment, such as dense trees or rapidly changing weather. Researchers hope that this form of AI can soon be used as the software behind self-driving cars, increasing reliability and safety on high-stakes tasks like driving on highways.
“Because these machine-learning systems are able to perform reasoning in a causal way, we can know and point out how they function and make decisions. This is essential for safety-critical applications,” says co-lead author Ramin Hasani, a postdoc in the Computer Science and Artificial Intelligence Laboratory (CSAIL). Their research will be displayed at the 2021 Conference on Neural Information Processing Systems (NeurIPS) in December.
The specific type of neural networks, called “liquid” neural networks, learn to complete a task through trial and error and change their equations to continuously adapt to new inputs. Thus, it creates a flowing, or liquid, form of calculation. The research relies upon a previous discovery, a brain-inspired deep learning system called a Neural Circuit Policy (NCP), which is entirely built by liquid neural network cells. The NCP is capable of autonomously controlling a self-driving vehicle with only a network of 19 control neurons.
Researchers observed that the NCP’s kept its attention on the horizon and borders of the road when making decisions while driving, which is the same way that a human would, and should, drive. However, other neural networks that the researchers had studied did not behave in the same way.
“That was a cool observation, but we didn’t quantify it. So, we wanted to find the mathematical principles of why and how these networks are able to capture the true causation of the data,” he says.
The reason, researchers say, is because the NCP learns to interact with its environment during a task and account for interventions, something that many AI cannot do. The NCP incorporates cause and effect into its decisions, making it one of the smartest programs to drive a vehicle. This new discovery can help increase everyone’s safety worldwide in the future.