MIT uses 19 neurons to achieve autonomous driving control

How many neurons are needed to build an autonomous driving control system?

Scientists at MIT tell you that the minimum number is 19! The method is to learn from elementary organisms such as nematodes.

Recently, a team from MIT CSAIL, Vienna University of Technology, and the Austrian Institute of Technology has developed a new AI system based on the brain of nematodes. The research results were published in the recent journal Nature Machine Intelligence.

They found that a single algorithm with 19 control neurons, connecting 32 packaged input features to outputs via 253 synapses, could learn to map high-dimensional inputs to manipulation commands.

This new AI system uses a small number of artificial neurons to control the vehicle's steering. The same autonomous driving system based on LSTM neural networks has a much more complex network structure.

△The vehicle control system implemented

Ramin Hasani, a postdoctoral fellow at MIT CSAIL and co-first author of the paper, explained why such a small number of neurons was possible:

The processing of signals within each unit follows different mathematical principles compared to previous deep learning models.

Inspired by the organism Caenorhabditis elegans, they proposed a neuronal circuit policy (NCP) in 2018.

The NCP approach is to repurpose the power of biological neural circuit models to create interpretable control agents and manage virtual and real-world reinforcement learning (RL) testbeds.

This method models the TW neural circuit of nematodes, which is mainly responsible for the reflex response of nematodes to external mechanical touch stimulation, and uses its synaptic and neuronal parameters as a strategy to control basic reinforcement learning tasks.

To test the new mathematical model, the team chose a particularly important test task - keeping a self-driving car in its lane. The neural network receives images of the road from a camera and then automatically decides whether to turn left or right.

Complex tasks such as autonomous driving often require deep learning models with millions of parameters. However, the NCP method can reduce the network size by two orders of magnitude.

This 19-neuron minimalist autonomous driving system uses only 75,000 training parameters, which is 2 orders of magnitude less.

The autonomous driving system built by the NCP method also needs it, but it is only used for the visual data passed in from the camera and extracts structural features from it. It has nothing to do with the actual steering of the vehicle.

The neural network behind it determines which parts of the camera image are important and then passes the signal to the network's NCP control system.

The control portion of the system, which converts data from the perception system into steering commands, contains just 19 neurons.

The two subsystems were stacked together and trained simultaneously on a dataset of videos of people driving cars in the Boston area, including data correlating images with the car’s steering maneuvers.

These are fed together into the network until the system learns to automatically connect images with the appropriate steering system and can handle new situations independently.

In addition to its simple structure, the autonomous driving system designed with NCP has two major advantages over traditional models: explainability and robustness.

The interpretability of the system allows us to see what the network is focusing its attention on.

As can be seen in the video, the neural network focuses on very specific parts of the image, such as the roadside and the horizon. The researchers say this behavior is unique among AI systems.

In addition, the interpretability is down to the level of each neuron. We can also see which neuron (lit up in the video) plays a role in driving decisions. We can understand the function of individual neurons and their behavior.

In order to test the robustness of the NCP model compared with traditional models, the researchers also added perturbations to the input images and evaluated the agent's ability to handle noise. As a result, NCP showed strong resistance to input artifacts.

In addition to interpretability and robustness, the NCP model has other advantages, such as reducing training time and reducing uncertainty in implementing AI in relatively simple systems.

Dr. Ramin Hasani also said that NCP can not only be used in autonomous driving, its ability to imitate learning means wider applications, such as automated robots in warehouses, etc.