Which AI is more human?
Created on Dec. 18, 2013, 4:11 p.m.
Disclaimer: I'm not a historian, or a student of AI, so please forgive any inaccuracies in my historic retellings...
Every now and again a article comes floating around the internet lamenting about the days of Good Old-Fashioned AI research, when men were men, and researching AI was like the adventures of a computer science version of Indiana Jones.
Good old fashioned AI was about using logic and reasoning to produce an intelligence. The approach was to make high level observations about a particular task or problem, and formalizing these in a logical and consistent way that would allow a computer to solve the same task when the situation varied.
The ultimate goal of this research was to create a machine with human-like logic and reasoning skills. This machine could then be given semantic knowledge about something or other, and use its skills to divulge solutions to new and unique problems.
But in the 1970s the AI Winter hit. This method of AI research, which had never claimed to be easy, had stagnated, and failures had accumulated. Funding was cut and for a long time it looked like AI was dead.
A new approach to AI appeared slowly gained traction. It was based upon statistical methods, and learnt from data, processing banks of information to try and make intelligent decisions.
Rather than reasoning, these methods appeared at face value to use mathematical tricks and brute force to get results. To the researchers from the good old days it was all artificial, and no intelligence. The new approach to AI was cold, mechanical. It was coined machine learning and damned by many for not being human.
This was all shown in the research of natural language. Early on great progress was made by Noam Chomsky and many other researchers in the understanding of the logical and systematic rules that encode natural and artificial languages. These opened large vistas of understanding and research in many other fields as well as language, but eventually their practically for real applications such as machine translation reached a stopping point.
It was so bad that IBM Researcher Frederick Jelinek became famous for his often quoted statement "Every time I fire a linguist, the performance of the speech recognizer goes up".
Natural language, with all its irregularities, was just not possible to encode in a handful of rules. Every slight variation and peculiarity broke these methods. The results were simply not of high enough quality to progress with processing times ballooning.
These days Google traverses the web and builds huge statistical models of language it uses to do its translation. All logic and reason is left at the wayside, and even the known and infallible structures of language are ignored. See: How I learnt to stop worrying and love statistics.
If human intelligence really is characterized by logic and reason it would entail that I could teach someone a foreign language simply by passing them a sheet with the grammar on it, and a list of words with their meanings.
Providing our brains were the logical reasoning machines, as assumed by early AI research, this should be sufficient to teach me something.
Of course this is not true. Even if I know you are not lying you cannot teach me something by just telling me it. While some humans are good at logic, none of us are good enough to build those kinds of connections in our brain without fretting.
On the other hand all of us are excellent at pattern matching. This means the inverse approach to teaching is almost always better. It is better to give someone a lot of examples, and to watch as they divulge the logical rules that govern the system themselves.
This isn't specific to language. Even in an extremely logical domain such as mathematics the most effective teachers teach examples first and the generalizations after. Talking about generalizations first and being specific later is an easy way to confuse your students.
Humans appear to learn the same way machines do.
Imagine we are trying to write an AI which can distinguish images of apples from oranges. In machine learning we give it some training data with correct answers and from this it learns how to classify new fruit.
If this system encounters apples 75% in training it will bias its classification toward apples, and only pick oranges if it is really sure of the result. This is called a prior probability.
Think for a second how horrible this idea sounds to good old fashioned AI researchers. The information as to if it is an apple or and orange is fully encoded in the image! The decision as to if it is an apple or an orange has nothing to do with what this system has encountered before!
The idea of probability as a whole appears horrible to these researchers. An thing is not 60% an apple and 40% an orange. It is either one or the other.
But prior probabilities have become are cornerstone of machine learning methods and are only disregarded at all costs by researchers. There is very good reasoning for this. Is because prior probabilities, like many other aspects of machine learning, are very human in their origin.
Prior probability is linked to the human concept of experience. Which is embodied in the storytellers guidance show, don't tell.
Many stories can be boiled down to a couple of sentences that state the point the story is trying to make. For example love conquers all or greed doesn't pay. But simply telling someone this sentence doesn't provide almost nearly the same impact as getting them to read a novel about it. Nor is it nearly as enjoyable, or accurate.
A novel adds weight and experience to certain situations. As humans we are naturally reserved creatures and we need engagement and evidence to believe something is true. Someone cannot just tell us it.
This is like the prior probability in statistical models. It makes us wary of things we have not seen before and gets us to hedge our bets when other factors make us uncertain. The real world requires scientific analysis to understand. Is it any surprise that machines should need to do this too?
The new AI has moved the logic from our conscious mind, to our subconscious and biological minds. It has rephrased the question from how to do we do this? to how to we learn to do this?. Perhaps in some people's minds we have still made a machine that is dumber before. If this is the case, why can it do so much more!
Not only does the new approach perform much better at many tasks, but it is arguable more human too. This asks a number of questions.
Are we really intelligent? Are we really in control? Are we really logical? I know what I think...