Could Neuro-Symbolic AI Change the Landscape of Autonomous Driving?

Automated braking. Adaptive cruise control. Driverless navigation. Most of us have felt the advancements in the transportation industry made possible through artificial intelligence (AI). Traffic management components are no exception. AI is already behind many of the advancements visible throughout our cities from automated traffic signals to high-resolution cameras. This technology is moving forward at a dizzying pace and is improving traffic safety tremendously along the way. Eventually, the cars on our roads will be replaced by autonomous vehicles (AVs), facilitating more optimal traffic conditions and less gas consumption. But let’s take a step back to consider one major obstacle that is still afoot – the capacity of AI to make inferences and use deductive reasoning. Before AVs can reach a point where no human intervention is necessary, our AI may first need to think more like a human.

Many of us have played peek-a-boo with a child. We hide our face with our hands, then exclaim “Peek-a-boo!” as we uncover our faces. There we are! Even though we have momentarily disappeared from sight, the child is able to discern that we are still nearby. This phenomenon is known by psychologists as object permanence and refers to the ability to recognize that an object still exists, even if it is not directly in one’s line of sight. Unlike a nine-month-old child, autonomous vehicles (AVs) are not yet at this level of reasoning. According to the Economist, “Autonomous vehicles are getting better, but they still don’t understand the world in the way that a human being does. For a self-driving car, a bicycle that is momentarily hidden by a passing van is a bicycle that has ceased to exist.” In other words, AVs do not yet have the capacity to grasp object permanence – a difficult task to train a computer.

There exists a potential solution to this dilemma: a symbolic AI approach. Symbolic AI is a branch of AI that has been around for a long time and is arguably more abstract than the typical data-driven methods we encounter today. This AI approach mimics human cognitive representations, where humans associate things or concepts with symbols which have structures and rules. Symbolic AI involves embedding behavior rules into computer programs so that a computer can transfer learning from one task into another, draw conclusions, use reasoning, and even make inferences.

There are few examples of applied symbolic AI to date, but imagine if a computer could make inferences or deduce reasoning, the applications would be endless. For a self-driven car, applying symbolic AI technology could even mean the difference between life and death. Say that an AV detects a cyclist riding along its side, however the cyclist temporarily disappears from the vehicle’s sensors. Rather than exclude the cyclist from its knowledge-base of surroundings, reasoning-enhanced software would consider the possible trajectory in which the cyclist may reappear, and take steps to avoid it if necessary.

Why has there been resistance to the true adoption of symbolic AI? One reason may be the elaborate coding required to build a system capable of making human-like connections and inferences. A system would need to have a knowledge-base capable of accounting for the varying factors that occur in the real world, which is tricky. Some may even say impossible. “You can’t define rules for the messy data that exists in the real world. For instance, how can you define the rules for a self-driving car to detect all the different pedestrians it might face?,” explains Software Engineer, Ben Dickson. In order for an AV to be ready to face the multitude of trajectories that may unfold on the roads, one must first create a massive collection of information.

Presently, a neural network-based approach is more frequently utilized in the world of AI. Using this method, a system is fed data and learns to recognize objects, patterns, and changes. For example, a computer is fed images of the roadway and it begins to recognize that all cars are traveling in the same direction. When the computer identifies a vehicle traveling in the opposite direction, it will understand this as an anomaly. Many AVs currently rely on this same technology. While this approach is effective for recognition and classification, there are drawbacks to its ability to predict movement and anticipate behaviors accurately.

Enter Neuro-symbolic Hybrid System

Some researchers and engineers have proposed a potential solution to fill the gaps between symbolic AI and modern neural network methodologies. Neuro-symbolic artificial intelligence, or NSAI, attempts to blend the two methodologies, taking advantage of the positives from each –

“NSAI is a hybrid between DL [deep learning] and symbolic approaches which attempts to capture the strengths of both fields. Deep learning has proven singularly successful in extracting complex features from data in tasks such as object detection and natural language processing. At the same time, symbolic AI is good for formalizing human-like reasoning. The objective of NSAI is extract features from data using DL approaches, then manipulate these features using symbolic approaches,” (Susskind et al., 2021)

A community of researchers from Harvard and MIT-IBM Watson AI have published a detailed study of this approach. They experimented with a video dataset called CLEVRER, standing for CoLlision Events for Video REpresentation and Reasoning. The short video showed different colored objects colliding. When applying a combination of neural networks and symbolic logic, the program performed relatively successfully in parsing through the video to understand the cause of the collision, predicting what will happen next, and understanding what would have happened if the circumstances were varied.

Research and experimentation with neural-symbolic AI methods over the last few years show promising advancements in the ability for AI to carry out reasoning. Now is the time for automakers to begin accelerating their research into AI methodologies. As a driving force behind progress in the AV field, it is up to the larger players to consider new AI approaches as they push for groundbreaking innovation.


Matt Hill is the Senior Data Architect at Rekor

Headshot of Matt Hill from Rekor