The “Intelligence” of a car is under scrutiny in these weeks after the accident where a self driving car hit, and killed, a woman that was crossing the road.
In a way people were surprised: how could a car that is able to navigate the maze of a city roads, avoid other cars be so stupid to hit a woman crossing the road? Never mind that the latest accident analyses by the police seems to indicate the car was not to blame for the crash, we would expect a car to be able to break instantaneously! It might not succeed to stop in time (the woman apparently stepped out from the side of the street to cross exactly when the car drove in leaving to time to break) but at least it should try. I, as a human driver, can have reaction times that are longer than the time the person drift from the side of the road to the front of my car or I might get paralysed by fear, but a computer with wheels should have no time lag and no emotion whatsoever.
Besides… do we need any intelligence in deciding what to do? There is an obstacle just break, for God sake!
Yet, there is plenty of intelligence that is required “before” taking the decision to break and this is where the issues at that crash probably arose.
In a recent statement the Waymo CEO (the Google spin off company aiming at placing self driving cars on the market) said that the technology on board their cars (Lidar) would have seen the woman crossing but he also emphasised that he did not know if it was a problem of sensing the presence of the woman or it was about deciding to break.
The first step a car’s brain has to take is to convert the data that its sensors are gathering into a meaning, in this case: what I see is a person crossing the street. This requires an analyses of the image (the image may be resulting from a digital camera or from a radar pattern, like the one created by Lidar). A nice article explains how the brain of a car, based on convoluted neural networks -CNN, processes these data and learn from them. The various data generated by the sensors are normalised (so that they can be compared) and then processed through a CNN consisting of 27 million connections and processing 250,000 parameters.
With this kind of figures one could expect that the time required for processing is huge, but that is not the case. The processing is being done by an NVIDIA chip that is able to process billions of operation per second. The crucial aspect here is that this processing needs to recognise what is going on, separating what can be a flat image, e.g. a drawing on the road, from an object. Notice that accuracy is of paramount importance, erring on the safe side is not good, because there is no “safe side”. Imagine a car that mistakes a drawing on the asphalt with an object blocking the road. It would break as hard as it can. What about the car following it? A collision would be almost inevitable since the human driver following the self driving car would not expect it to brake out of the blue.
Possibly some may object that recognising an object is not requiring an intelligent process. What about evaluating the probability that a certain object may move and cross our path? That can be a cat, but it can also be a rock rolling down a slope…
What about recognising that an object is moving to intercept our path and also knowing (assuming) that it will become aware of our path and changing its path to avoid the collision? And what about guessing which way it will change its path?
As you can see, the more one thinks about the possibilities the more complex it becomes and it soon requires the development of a virtual model where different alternatives can be played out. You may not like to call it “intelligence” but this is what intelligence is about!
As a closing remark I note the ambivalence that we have towards the “machines”: on the one hand we like to consider them “stupid”, they are not like us, aren’t they!? Hence they are stupid! On the other hand we expect the machines to be perfect, we can get it wrong, we can get confused… A machine is not supposed to become emotional, to be distracted, to be confused, a machine is supposed to work like a .. clockwork!
There seems to be a general agreement that in the Uber crash the chances of a human driver avoiding the crash in that situation were “zero”. Yet, we expected the autonomous car to be “better” than a human driver (which it is normally the case) and because of that the reaction of most people was that autonomous cars are “dangerous”! This has nothing to do with reality, yet perception is everything.
There’s more: as we are creating more and more complex machines, more and more flexible as they have to be to face a changing unpredictable environment, they are also becoming more and more like us, for better or worse.