When my favourite chat app (used by myself and over a billion other people) figures out all by itself when to use autocomplete and when not to use it, then I will be impressed by machine learning, artificial intelligence, or whatever term you prefer to use. At this moment its performance is nowhere near what I consider to be genuinely intelligent, as it is even unable to figure out which language I am chatting in (although the language of the chat sessions typically depends on the person I am chatting with, so that should be a solid hint for a so-called intelligent system). Obviously I can switch off the autocomplete function to avoid words in one language being autocompleted by extremely unrelated (in terms of my desire to see them pop up there) words in another language, but in some situations – for example when the correct suggestion for a 20 character word pops up after having typed just 3 characters – I definitely am happy to use the autocomplete.
Sure, there have been some impressive performances based on machine learning – like IBM’s Watson beating the best human Jeopardy players or Google’s DeepMind winning at the board game Go. But those are basically one-trick-ponies. Those systems are extermely powerful in fulfilling one specific task, but more or less useless at most other tasks. A very striking comparison for them I have heard somewhere (I would attribute the author if I would have remembered who I heard it from) is that of a 9-year old showing off factual knowledge picked up from reading books, but lacking the necessary life experience to put it to any decent use. [On a personal note¹: I now realize I was that kind of kid at times – like that time (at age 9 or 10) I entered into a ‘European countries and their capital’-contest with the teacher, and won it because I knew what the capital of Liechtenstein was².]
Human intelligence on the other hand is often demonstrated by performing good to excellent at a vast array of things. And it always surprises me how difficult it is to explain (in terms of programming) to a computer some tasks that most people will consider relatively easy. As I am working a lot with databases, let me give you a specific example from that domain: in a database of online purchases, where the data are filled by the customers, you will inevitably end up with some records where the person’s first name is filled in the field that is supposed to contain the last name and vice versa. Online shoppers’ main priority is getting hold of whatever they are looking for, not database hygiene. For humans most of those first name/last name switches are quite easy to spot, but the effort it takes to create a computer algorithm that will perform at a comparable level is usually much greater than simply having a human do this manually – unless the database is huge of course.
Does that mean that I think a really intelligent Watson will never exist? Of course not, but I also don’t think it is really around the corner. Huge progress is being made in artificial intelligence and machine learning; fascinating things are happening in domains such as natural language processing; and I welcome breakthroughs in all of those fields. Today, however, I am simlpy not yet impressed. But then again for the average human it also takes quite a few years to reach that level of being good to excellent that vast array lot of things.
¹ Yes, I know it makes no point to add something like ‘on a personal note’ in a personal blog.
² No, I’m not going to tell you what it is – looking it up now will be a much smaller effort for you than it was for me at the time (we’re talking 1970’s here).
Pingback: Machine learning in everday life | THE SECOND HALF
Pingback: OK, this time I am impressed | THE SECOND HALF