A large rubber ball bounces into the street, the driver sees the ball, intuits that a child might follow, and stops the car. Automated vehicles are incapable of accomplishing the task of determining what pedestrians are going to do.1 Humans intuitively develop a theory about how people behave.
We do the same when we view a representative painting. We use our previous experiences, and our physical reactions, to imagine a world of possibilities. Viewing the Vermeer painting, The Love Letter, we look through a doorway surrounded by drapes. Our perception combined with our past experiences gives us the idea that we are looking into a private scene. We feel as if we are intruding, and this makes us wonder what are we seeing.
Is that a worried glance on the woman who has interrupted her lute playing to get a letter? We ask ourselves why did she stop to take the letter, and why does she look the way she does? Is it worry, or is it concern, or is it anxiety, or anxiousness? These reflect on our experiences of these feelings and emotions, and based on our memory of our reactions, we compare those with that of the woman.
We know what love letters and lutes are. We know about the world. Seeing the world, having visceral experiences impacts the mind. Our mind then interprets our experiences, and we then re-experience what we see, and that changes the impact on the mind. We use our knowledge about other people to try to decipher what the facial expressions, bodily actions mean based on our experiences. And so it goes, as we continue to see or feel new things about the work of art, which changes our mind, and then changes our experiences for as long as we want to observe and analyze the work of art.
These are reciprocal interactions of brain and body. To do this we have to make decisions about the world. Thoughts and feelings contribute equally to these choices. We use both interoception, the ability to understand the sensations arising within our bodies, and abductive reasoning which tries to find the most likely explanation for a set of observations. It looks for a plausible conclusion, even if we have never seen that situation before.
No computer can do this today. Today’s artificial intelligence, including that in automated vehicles, is about pattern matching of signals received from sensors and using inductive reasoning. Automated cars do not reason and learn. It cannot figure out that the bouncing ball might be followed by a child unless it has been programmed to do that. Would it recognize the same situation if an America football came bounding out, or a Spaldeen, or Pensie Pinkie?
Whether we experience a work of art, or figure out what a pedestrian, or cyclist, or motorist is going to do, our mind is not a computer instructing the body what to experience. The mind is a biological organ with a biological body, not a disembodied thinking machine. Alan Jasanoff, among others, has argued that the metaphor that our brain is a central computer has outlived its usefulness. Continuing to think about the mind or the brain in this way is harmful. Both art and self-driving cars require the same experience.
Basing the automated car’s intelligence on a neural network does not allow for how we live as humans. Whether we are looking at art, or driving a car, we can experience the world, and generate a model of it that will allow us to endow it with meaning, which will allow us to anticipate what might happen, or what might happen next. This anticipation is often exactly what creates surprise in art, such as the famous C major chord in Haydn’s Creation. 2 Just like with art, a self-driving car has to be prepared for the unexpected.
Is a pedestrian going to step into the street, or stay on the sidewalk. Is the running pedestrian going to stop or enter into the middle of the road. Who knows what the weaving cyclist is going to do? Run the red light? Often the pedestrian and the driver look at each other, and somehow figure out what to do.3
My argument is that besides being a dynamic control task, driving, like art, is a social phenomenon and requires interactions between all road users involved to ensure the flow of traffic, and to guarantee the safety of others. Recent studies of autonomous vehicles show how the lack of social understanding can result in traffic accidents or erratic behaviors towards pedestrians. 4 There are local variations that make this even more complex, such as the illegal Pittsburgh left:5 Despite being illegal, it is common practice nonetheless.
Many years ago Massachusetts changed its laws about who gets the right of way in rotaries, and who has to stop at stop signs. How will the automated car change when the law changes, and there has been no change in the training. 6
For cars to be truly automated, they need to abandon the computer metaphor, and treat intelligence as embodied. Our brains are biological organs embedded in a biological body. Our minds are not one way control centers over the body. Our minds are not computers. The world affects the brain, and feelings arise, interpretations arise, which then change our view of the world, which gives rise to new feelings and interpretations. To do that we need a “theory of mind” as the neuroscientists call it, to figure out what the pedestrian is going to do. In other words we put our selves in their place and given what they are doing, we anticipate what they are going to do. But a theory of mind is not enough, we need to understand how that mind is embedded in a body. What neuroscientists call “embodied cognition“. Art makes no sense without this embodied intelligence. Hence, self-driving cars need this relationship with the outside world, thinking about art and self-driving cars clarifies this concept.
Right now the best automated vehicles can do is to react to what they have been taught – either by programming, or machine learning. They cannot make a hypothesis and project – ball in the street — child may come out. Even if taught about a soccer ball, what about a football bouncing into the street? That is more than processing signals at the edge.7
How can you look at a portrait without reading into the facial features and pose, based on your understanding of the world? You cannot separate the brain from its environment. You are able to look at people in the street, on a subway, and can construct a story about them from their faces, body positions, or actions. That is embedded cognition with a theory of mind. Looking at a landscape painting requires you to immerse your brain into imagining yourself (even if you do this instinctively, and not consciously, or deliberately) looking at a landscape. Looking at a work with religious or spiritual value, you have to imagine how this work embodied the values and beliefs of the creators or the worshipers .
Even the disembodied brains of the Star Trek episode, “The Gamesters of Triskelion” could not be just satisfied with their thoughts, they could not be Aristotle’s unmoved thinker, they need stimulation from the world around them.8
How could our current generation of computers understand the awe and splendor of a Yayoi Kusama Infinity Room? My experience of infinity was felt throughout my body, and it was my brain that had to sort those feelings out. My brain did not tell my body what to feel. As the brain processed those feelings, of course, the experience was refined. It was an embodied brain, not a command and control center that had the experience. An automated vehicle will only be truly possible when the relationship between art and self-driving cars is not bizarre.
- Work is only being done on simulations. See https://arxiv.org/pdf/1805.11773.pdf
- position 7:15-7:50 in the recording.
- All this is active research, nobody has any real answers yet. See:
https://www.ifsttar.fr/en/online-resources/science-and-society/transport-and-mobility/science-topics/automated-vehicles/interactions-with-pedestrians/
https://arxiv.org/pdf/1702.00785.pdf
https://eshed1.github.io/papers/humansTIV.pdf
- https://slate.com/technology/2016/03/google-self-driving-cars-lack-a-humans-intuition-for-what-other-drivers-will-do.html https://www.nytimes.com/2015/09/02/technology/personaltech/google-says-its-not-the-driverless-cars-fault-its-other-drivers.html
- https://en.wikipedia.org/wiki/Pittsburgh_left
- In Massachusetts rotaries, the car entering the rotary used to have the right of way; that was changed later to give the right of way to the cars already in the rotary. Massachusetts used to have the rolling stop rule. If three cars are lined up at a stop sign, only every third car (starting with the first) had to completely stop.
- https://www.infoq.com/news/2019/02/Grady-Booch-Future-AI
- https://en.wikipedia.org/wiki/The_Gamesters_of_Triskelion
Interesting as usual, Michael–I’m always happy to see a new post from you. But suppose you’re right (although I’m not wholly convinced about this). Even if there are some things that self-driving cars won’t do as well as human drivers, isn’t it possible that overall driving safety could still improve? There are many, many common human mistakes that self-driving cars will never make, chief among them driving while intoxicated. Even if a car can’t intuit the meaning of a ball bouncing into the street, I think I’d be willing to make this trade-off.
I think you misunderstand my point.
It is not that self-driving cars will not someday be possibly better than humans, the point is that the current generation of self-driving cars never will be.
The reason for this is they are based on an incorrect assumption about what intelligence is. You cannot program a car to cover all situations, and supervised learning without the ability to independently learn from experience, and deal with situations not exactly seen before, will never gives us cars that can behave satisfactorily in all the situations humans do. The ball bouncing example, the Pittsburgh left, the change in Massachusetts laws, are just simple examples of an enormous number of situations that are not at all improbable. They may be unlikely, but they occur often enough. The more self-driving cars on the road, the more likely these situations will occur.
There is not enough training data, you need vehicles that can draw conclusions from experience, and hypothesize about unknown situations. Nobody really understands unsupervised learning.
The current generation of automated vehicles may not get intoxicated, but they do not even have common sense. The best we will have is the so-called, Level 2 or 3 driving. For example, interstate trucks driving under the supervision of a driver actively monitoring its behavior on well-marked highways. The driver takes over in the city.