Full width home advertisement

Welcome Home


Post Page Advertisement [Top]

The Development of Artificial Intelligence That Is Capable of Recognizing Cause and Effect



Neuronal networks are capable of learning to solve a wide range of problems, ranging from recognizing cats in photographs to steering a self-driving car, and they are being used to do so. Nevertheless, it is still unclear to what extent these powerful pattern-recognition algorithms comprehend the tasks that they are tasked with accomplishing.


If a neural network is tasked with keeping a self-driving car in its lane, it may learn to do so by observing the bushes along the road's side, rather than by learning to detect lanes and focusing on the road's horizon as is currently the case.


Research conducted at the Massachusetts Institute of Technology has demonstrated that a specific type of neural network is capable of learning the true cause-and-effect structure of the navigation task for which it is being trained. These networks should be more effective than other neural networks at navigating in a complex environment, such as one with dense trees or rapidly changing weather conditions, because they can learn the task from visual data.


As a result of this research, machine learning agents performing high-stakes tasks, such as driving an autonomous vehicle on a busy highway, have the potential to become more reliable and trustworthy in the future.


As a result of their ability to engage in causal reasoning, we can deduce and explain how these machine-learning systems operate and make decisions." According to Ramin Hasani, a postdoctoral researcher in the Computer Science and Artificial Intelligence Laboratory and co-lead author on the paper, "this is critical for safety-critical applications" (CSAIL).


Co-authors include Charles Vorbach, a graduate student in electrical engineering and computer science at CSAIL; Alexander Amini, a PhD student at CSAIL; Mathias Lechner, a graduate student at the Institute of Science and Technology Austria; and Daniela Rus, the Andrew and Erna Viterbi Professor of Electrical Engineering and Computer Science at CSAIL and director of the Center for Advanced Information and Communications Technology. During the Neural Information Processing Systems (NeurIPS) Conference in December 2021, the findings of the research will be presented.


An example of a machine learning technique is neural networks. In this technique, the computer learns how to perform a task through trial and error after studying a large number of training examples. Additionally, "liquid" neural networks adapt to new inputs by altering the equations that govern their operation.


Based on previous work, Hasani and colleagues demonstrated how a brain-inspired type of deep learning system known as a Neural Circuit Policy (NCP) composed of liquid neural network cells is capable of autonomously controlling a self-driving vehicle with only 19 control neurons in a previous paper.


When the NCPs were performing a lane-keeping task, the researchers observed that they maintained their attention on the road's horizon and borders when making driving decisions, just as a human would (or should) when driving a car, the researchers concluded. Other neural networks that they looked at were not always oriented toward the road, as they discovered.


"It was an interesting observation, but we didn't make any attempt to put it into numbers." Therefore, we sought to understand the mathematical principles underlying why and how these networks are able to capture true causation in data, as explained by Dr. Cheng.


After being trained to perform a task, they discovered that an NCP gains the capability of interacting and accounting for its environment as well as accounting for interventions. In essence, the network determines whether or not an intervention has an effect on its output and then establishes a link between the cause and the effect.


A forward run of the network is used to generate an output, followed by a backward run to correct for errors that occurred during the training. Scientists have discovered that NCPs can relate cause and effect in both forward and backward directions, allowing the network to concentrate intensely on the true causal structure of the task at hand.


Hasani and his colleagues discovered this causality without imposing any additional constraints on the system or putting it through any special configuration.


The ability to characterize causality in safety-critical applications such as flight, says Rus, is critical in these situations. We investigate the causality properties of Neural Circuit Policies for decision-making in flight, which includes flying in densely forested environments and in formation, as part of our research.


They assessed NCPs through a series of simulations in which autonomous drones were used to perform navigational tasks on the battlefield. Each drone navigated by taking cues from a single camera for guidance.


To accomplish this, the drones were instructed to travel toward a target object, chase a moving target, or follow a series of markers in a variety of settings, including a redwood forest and a neighborhood. The group also encountered a variety of weather conditions, including clear skies, torrential rain, and fog during their journey.


However, the researchers discovered that while the NCPs performed similarly to the other networks during good weather, they outperformed them all when faced with more difficult tasks, such as running after a moving object in a rainstorm.


According to what we discovered, NCPs are the only networks that pay attention to the object of interest while completing the navigation task in various environments, as well as in a variety of lighting and environmental conditions." The only system capable of doing so casually and actually learning the behavior we intend," he says, is the one he developed.


Furthermore, the researchers' findings suggest that the use of NCPs may also allow autonomous drones to navigate successfully in changing environments, such as a sunny landscape that suddenly becomes foggy.


The system can perform admirably in a variety of scenarios and environmental conditions once it has figured out what it is supposed to do, says the researcher. This represents a significant challenge for non-causal machine learning systems currently in use. This research, we believe, is particularly interesting because it demonstrates that causation can emerge from the selection of neurons in a neural network," the author explains.


The researchers intend to continue their investigation into the use of NCPs in the construction of larger systems in the upcoming years. This means that they may be able to handle even more complicated tasks by connecting thousands or millions of networks.

No comments:

Post a Comment

Bottom Ad [Post Page]