Features

net.wars: What movie was that, anyway?

This week, Wendy has clearly been searching for a lost movie, and finds computers no help at all ...

The late Roger Needham commented a number of times that someday we might be able to stand in front of a video library and say, "Show me the one with the red-haired actress who gallops off on a white horse at the bottom right of the screen" and have the library respond with the requisite scene. But not yet.

Of course, there a lot of humans who can't do that either. We'll except Martin Scorsese - he's said to have an encyclopedic memory for the details of all the films he's ever seen. But the comment got me thinking, not so much about the limitations of information retrieval, Needham's specialist area, but about how inaccurately "memory" describes what computers do.

Computer memory as we currently know it works nothing like human memory. The analogy is really very poor.

For one thing, it is not predictable what will open the door to a human memory. Humans recall things using any combination of the five senses. Computers have either no senses or only one and they don't experience the world directly: whether it's sound, visual input, or keystrokes, everything comes down to a set of files linearly ordered. Computers do not recall memories; they retrieve and replay data.

When you want to remind a forgetful human, you feed that person more detail. "Yeah, you do. Remember, it starred Katharine Hepburn and Peter O'Toole, and they were spending Christmas in this castle fighting with their sons, and it had that great trumpet music. Oh, come on, we went to see that when were 17 at the Thalia, and you were wearing that awful brown coat I always hated, and you promised me that night you'd throw it away and you never did."

The human now has several different ways to access the memory of the movie: the catalogues of movies seen at that particular theatre, discussions about the brown coat, images of Katharine Hepburn, Peter O'Toole, horses, and castles, trumpet music, even things you remember doing with that particular companion when you were 17.

But even that number of avenues to the memory doesn't capture the complexity. Are you a sound person or a visual person? Do you remember the image of Katharine Hepburn staring at herself in a hand mirror, or the sound of her voice saying, "How, from where we started, did we ever reach this Christmas?" Or is your memory best jogged by the smell of popcorn?

This is not how computers remember things. If you're lucky, and you've planned in advance, the file you're looking for will have been given a file name that makes it easy to find: lion_in_winter.avi, perhaps. If you're a little less lucky, the file will have been given keywords you can use to search for it with: Hepburn, O'Toole, Goldman, "Lion in Winter". If you're not lucky at all, the file will be called 19681000541.avi, and it will be sitting in a directory with 9,842,637 similarly numbered files and you'll have to open all of them to find out which it is. This is not the computer remembering things for you; this is you remembering things for the computer.

As the chapter on learning and memory in Cognitive Neuroscience: The Biology of the Mind, makes plain, there are many types of memory that have no computer analogue. Procedural memory, for example, lets us remember how to perform actions like typing or riding a bicycle without having to think about them in detail. And you can bolster human memory by using multiple senses - it's easier to remember a new phone number if you both say it out loud and write it down.

But the other important thing about human memory is that often it's an act of communication. In the example above, I jog my friend's clearly failing memory (at least, it's failing until I say, "Oh, god, that wasn't you, that was your friend Steve") from many different directions because even when you're very close you don't necessarily know which bit of shared experience will loosen the chunk you want. With a computer, you can only try different lines of approach to locating the file, such as remembering where you might have put it, how you structured your directories at the time, whether it was this computer or the old one, whether you might have stored it on a CD, whether you gave someone else a copy, or where you got it from and could conceivably get another copy. That is not how humans jointly remember; that is how humans look for a lost shoe.

Computers don't even really have the advantage in bolting on capacity - humans build new neural pathways as part of a process called consolidation.

The one place the computer-human analogy is useful is when you want to help a human understand how a computer works in general terms. RAM=short-term memory; hard disk=long-term memory. The problem, however, is that increasingly humans are using computer terms to explain how their own memories work.

We've always assumed eventually we would reach artificial intelligence by making computers "smarter" - making them more powerful, giving them greater capacity to record data, increasing the types of data they can search effectively, and trying to find clever ways for them to make human-like connections. But perhaps what will actually happen is that we'll reach artificial intelligence by constraining our own abilities to match those of computers.

PS: I still have no idea what, if any, movie Needham had in mind.

Thanks to Terry Hines for his assistance.