Artificial intelligence faces serious roadblocks
Over-hyping technological advances can obscure the facts of what computers can actually do, and what they can't – driverless cars are a case in point
The wonders of technology promotion never cease to amaze. The trick is to catch the attention of the press and the curiosity of the public and not get pinned down by details.
Self-driven cars are one of the current technologies getting a great deal of public attention as a game-changing development, hailed by automobile manufacturers responding with enthusiasm.
These promotions rely on public ignorance of what the new technology can actually do and what it can’t. This allows the imagination to run free about its future benefits. The beneficiaries of the hype are promoters of a brand or company, or ambitious entrepreneurs looking for the next big thing to raise money.
In this case, for example, Google benefits from the publicity it gets as an innovator. And Uber can dangle the future of driverless (hence cheaper) transportation and therefore a more profitable business.
The most prominent area of ignorance is the public perception of what computers actually do relative to human activities, and in particular human thinking.
As defined by Wikipedia, “thinking allows humans to make sense of, interpret, represent or model the world they experience and to make predictions about the world”. In the case of driving a car, for example, the driver responds to traffic conditions based on reactions that are conditioned by previous experience and calculation of least risk moves. This is why driver training is important and why drunk drivers crash, since they are incapable of bringing best-experience data stored in the brain to address the actions needed for the immediate conditions.
Now, suppose a computer is driving the car. In effect we need to replace the human thinking with the “intelligence” of the computer. The computer has none independent of what it has been programmed to do. So in human terms, it is like having an idiot where every move is determined by a human contact.
The difference between a human idiot and a computer is that the computer stores information and reacts as programmed, while an unsupervised idiot is unpredictable.
So why the excitement? Because we hear of a technological answer, something called “artificial intelligence” and “machine learning”, whereby computers are programmed to overcome their fundamental “thinking” limitations compared with humans.
What has been done since the invention of computers is to program them to respond to defined situations with defined actions. So a computer will appear intelligent and even superior to humans in a defined situation with no surprises. But such software does not make the computer act appropriately when faced with totally unexpected situations.
Here, programs can be developed using inference factors so that perceived situations are matched on a probability basis with stored ones and actions ensue on that basis. If no match occurs, the computer is at a loss or might act on the next most probable stored event.
In any case, what the computer does is to determine an action based on the most probable stored event conditions. Obviously in a car, the number of situations requiring response will vary with the surroundings.
In a well-controlled environment (like driving on a track), the computer can be expected to respond to situations consistent with programmed information. The problematic situations are the accidental ones when something happens on the track that requires a quick response different from the programmed actions.
This is where the awareness and quick response of a human driver comes into play and where the response of a computer making the decisions is quite another matter. And this is the skill that differentiates race-car drivers from the rest of us – and computers from all of us.
The more potential events that a computer has to react to, the more difficult the ability to program becomes. Just imagine city traffic in most countries of the world where road conditions can physically change hourly, not to speak of traffic patterns.
These issues limiting truly self-driving cars are lost in the hyped hope for wonderful things like reducing traffic deaths, allowing incapacitated people to use cars, and saving on the cost of professional drivers.
Let’s say the driverless car is traveling down a two-lane street at 60km/h and an out-of-control car swerves and heads straight at you. Does the computer decide to hit the lamppost on your right or the baby carriage on your left? Do you broadside a bread truck or a school bus?
Self-driving computers require enormous detailed information about your route. Their sensors can’t recognize a pothole in front of you, or an oil slick, or black ice that appeared in the last hour. They can slam on the brakes if a person or vehicle appears in front of you. But they can’t make life-and-death decisions in unpredicted and unpredictable situations.
And finally, an obvious question: Is the average consumer actually interested in programming his car before every trip? Or being reliant on a computer connected wirelessly to the “cloud” with its potential reliability issues?
Sure, people now interface with computers constantly, but they are used to programming being well hidden from them – the smartphone is an example of getting action on the basis of pushing buttons. Being uninterested in the guts of the machine, the average person wants his computer interface to be intuitive and unsophisticated. The invisible hard technology is under the covers. How likely is it that this will be achieved with a mass-production automobile given the vast variety of global driving conditions?
Of course, computers can greatly improve the safety of cars and ease driving stress – but replace the steering wheel?