Auto & Transportation

A few days ago, according to foreign media reports, researchers from the Massachusetts Institute of Technology (MIT) demonstrated that when a certain neural network is trained to perform a navigation task, it can understand the true causal structure of the task. This research can improve the reliability and credibility of high-risk machine learning agents, such as driving self-driving cars on busy highways.

It is reported that this type of neural network can understand the task directly from visual data. This type of neural network is more efficient than other neural networks when navigating in complex environments such as places with dense trees or rapidly changing weather conditions. The new research utilizes previous research by Hasani and others. The latter showed how a brain-inspired deep learning system, Neural Circuit Policy (NCP), built by liquid neural network cells, can pass only A network of 19 control neurons automatically controls a self-driving car.

Researchers have observed that NPCs performing lane keeping tasks will focus their attention on the horizon and boundaries of the road when making driving decisions, which is the same as when humans drive a car, while other neural networks studied are not. The club will always focus on the road. They found that when an NCP is trained to complete a task, the neural network learns to interact with the environment and understand the intervention behavior. In essence, the network can identify whether its output has been changed by some kind of intervention, and then link cause and effect together.

During training, the network runs forward to generate output, and then returns to run to correct errors. Researchers have observed that NPC associates causality in forward and backward operation modes, allowing the network to focus on the true causal structure.

Researchers found that in good weather, NPC performs as well as other neural networks on simpler tasks, but on more challenging tasks, such as following moving objects in heavy rain, NPC performs better. For other neural networks. In the future, researchers hope to explore the use of NCP to build larger systems. Connect thousands of neural networks together, allowing them to handle more complex tasks.