The Next Frontier in AI — Replacing Statistical Independence with Intuition
It takes a human about 20–30 hours to learn how to drive a car, while it takes tens of thousands of hours to train a neural network to achieve this same capability. Even after all of these years of training and despite of using the latest and greatest in processing and sensor technology, self-driving cars are still not deemed road-safe.
World Model
When humans learn how to drive, they already have a basic understanding of the world. They have a basic intuition of how to recognize a dangerous situation and they can fill in the blanks under incomplete information. These abilities are critical in many life-or-death situations that may not occur frequently, but when they occur there would be significant negative impact on the driver and their surroundings. We can train our learning model to recognize many of these situations, but there is an infinite number of them and even after millions of miles driven, the machine learning model will not have experienced anywhere near all of them. Why? Because deep learning models do not have an inherent understanding of how the world works. They do not know any laws of physics and neither do they know ethics or even liability laws. Everything they learn is based on statistical independence of all input variables. Humans however operate by making implicit assumptions on how some of these input variables are correlated. The video shows a very basic situation of how neural networks learn by going through literally every iteration of a mistake. They even have to make these mistakes separately, when going left and right. This is statistical independence at its finest.
Human Intuition
When you see a ball rolling on the road, you automatically watch out for children playing. When you see a car with a bunch of mattresses precariously strapped to its roof you change lanes and ideally pass that car very quickly, or when you are trying to get out of your parking spot after soccer practice you know that unless you actually inch forward a little, nobody will let you out and you will be there forever. But how abruptly or how far should you inch out?. This depends on a lot of factors that can even be dependent on the individual. For example, you recognize your friend’s car and you know that they will let you out, so you are more confident backing out. Or the other car flashes its headlights and the driver turned her head and nodded at you, which is also a clear sign amongst us humans that you can safely proceed. While you could teach deep learning models to recognize many of these clues, you could not teach them all of them, without the learning model fundamentally understanding the world.
Statistical Independence
I deliberately picked this heading as this is the “next frontier” in deep learning. We can have people sit down and come up with long lists of real-life situations, e.g. in traffic, that can be taught to the machine. But the problem is that this just a finite list and that we can never capture all of the subtleties involved in human decision making. Often, humans do not even fully know themselves why they had this hunch that a kid might come out of nowhere on a skateboard to fly in front of their car. The human brain simply uses all of the sensory and cognition power it has available to optimize its decision making processes, while machines are limited to a finite number of input variables they have deemed as potentially relevant for this specific situation. They do not have the ability to transfer knowledge from previous seemingly unrelated situations to the current one. We might have listened to a friend talk about treacherous ice that led to a car wreck despite the car showing 40F on the display. As our friend told us more about the accident, we developed sensitivity for places where there could be ice despite overall temperatures suggesting otherwise. To teach our machine this same skill, it would either need to go through these very rare occurrences or we would need to hardcode these situation into its decision engine. Both options are not viable, as this would be an endless task with too many independent variables involved. In a nutshell, today’s bottleneck is the statistical independence of input variables.
Self-Supervised Learning and Statistical Independence
Self-supervised learning refers to the approach of feeding neural networks with lots of input materials that contains gaps for the model to gradually improve its ability to fill in those gaps. While this is an ingenious approach, the neural network learns in a brute-force manner that makes up for its lack of understanding of the world. Take the sentence: “The … fell off the … and broke a …”. Even if we have already filled in the first gap with “man”, the model will still try to fill out the other two gaps with all of the terms that cannot possibly fit. “The man… fell of the … mouse, aquarium, goldfish, fork, wall, socket, and so on, and broke a … hamster, goldfish, fork, chair, tree, and so on.” And even after the 1,000 example of a man falling off a ladder and breaking his leg, the model will still not have the intuition that “ladders lead to accidents that can involve broken bones.” Or maybe that the older the man and the wobblier the ladder, the higher the probability of him falling off and sustaining severe injuries. Taking this example further, the model will also not transfer this concept of ladders leading to accidents to situations that are very similar, such as myself precariously standing on a chair to reach the top shelf of my in-laws’ closet. Unless it sees me and lots of others fall down in this type of situation, it will not deem the concept of “standing on chairs” a risk factor.
And that’s why we do not have self-driving cars yet.
