The Blockchain, AI , and Self-Driving Cars: A Step Forward

What do you get when you mix the Blockchain with an AI system and a network of self-driving cars? Volvo, GM, Honda, Toyota, BMW, and several others are seriously invested in building the future of autonomous vehicles and yet, none of them seem to be fully confident in it at the same time.

Across the automobile industry, the consensus appears to be that AVs just aren’t safe. This opinion is helped along by scary cases of AVs losing control like Uber’s fatal accident involving one of its’ AVs in 2018 and Google’s less serious, but still meaningful crash in 2016.

On top of these two examples, one can only imagine how many others may have gone unreported to the press. Since this is just speculation, however, for now, it is more important to move ahead to what we do know. What is arguably most important to the continued progress of the AV space is that it continues to learn from any and all accidents that its’ vehicles experience.

Here, by learning, I mean taking in data continuously from the road and improving, instead of stagnating. If you’ve read any of my other posts on how AI networks are structured, then you know that as of now, this refers to using a supervised learning framework. At face value, the problem with this is that if a system runs on supervised learning-based algorithms, then by definition, it is not autonomous.

Therefore, the question then becomes: how AI teams can incentivize continuous, yet reliable and secure learning in the systems that run AVs?

To begin, AVs need to pivot to using unsupervised learning principles to analyze the data they take in. The problem with doing so is that not only is the unsupervised learning space the more underdeveloped niche, leaving an AI to learn on its’ own comes with its’ own issues.

Take the sub-field of unsupervised learning called reinforcement learning, for example. While many of the knowledgeable sources on the subject seem to be biased towards an optimistic view of its’ abilities, certain truths can still be gleaned from their efforts to educate the public. Before we jump into what some of these truths might be, it’s important to review the definition of reinforcement learning.

Without getting lost in the weeds, the easiest way to understand reinforcement learning is to think about how humans learn. Typically, we do something again and again until we learn the correct way to do it. This extends to speech, movement, and just about any action that it’s possible to take.

If you extend this general idea to autonomous vehicles, the same is generally true. Those that are structured on reinforcement learning principles, learn from failure. On a deeper level, this also means that each possible action of each car would need to be coded within a reinforcement learning framework.

If you’re wondering about what this means on a step by step basis, check out our link from the Towards Data Science blog below. For now, in the interest of clarity, we’ll stick to the main point of this discussion, which relates to an announcement from CoinDesk today.

Reportedly, GM and BMW are throwing their resources behind research related to using the blockchain with self-driving cars. Overarching this is an organization called the Mobility Open Blockchain Initiative. While, it’s not entirely clear what the project will entail at this time, according to CoinDesk, the MOBI members are working to create a blockchain-based platform to facilitate data-sharing between many AVS at once.

For a clearer picture of what this might be, think about what the Internet of Things is and then consider it at a smaller scale, for only one type of device.

With such a platform, the era of truly “learning” AVs could be ushered in, provided that enough companies are allowed to participate. Ideally, the platform could serve as an excellent example of how reinforcement learning can be used as an improvement mechanism for many devices at once, at scale. Still, it’s important to note that as of now, it is not exactly clear whether or not the car companies will include reinforcement learning in the project’s foundational framework at all.

What is more clear is that the companies involved want to push the training of AVs to a whole new level. Until now, most efforts related to testing AVs have been kept close to the breast. If these companies are successful, we might see truly massive growth in the space due to a collaboration, the likes of which we have never seen.

Data may no longer be the new oil.

AVs seem destined to go mainstream, but the question remains as to who will successfully develop the space for them to teach each other to drive safely. Until the MOBI effort goes live and is validated, there is therefore, no reason for any company to express full-confidence in AVs as they are now.

We need more data for that. In future posts, we’ll dig into why this is true on a more technical level.