For my birthday/Christmas—they’re essentially one and the same—I received what is perhaps the most wonderful card game ever. Apples to Apples is a group adventure in forced classification, and its parallels to data mining techniques are vivid, fun, and revealing. Its power is so subtle that my sister almost didn’t realize just how much of her bedroom practices she had revealed in a single card. My dad caught it, though. Let’s hold back a moment. You need to know how to play the game first.
There are two kinds of cards: red apple cards—each player holds seven of these at all times—and green apple cards, which stay in the middle. Both cards have a single word printed on each of them. The red cards have nouns like The Godfather or exorcism. The green cards get adjectives like scenic or cold. Like the name suggests, the gist of the game is to match Apple to Apple.
Each round one person reveals a green apple card. This person is the judge for the round. He gets to choose which person comes up with the best pair. Let’s pretend the word for this round is ‘flirtatious.’ Everyone else looks in her hand of seven red cards and picks the one which is the most flirtatious. Here’s a list of the seven cards in my hand:
- The Olympics
- Plane Crashes
- The Ozone Layer
- The Opera
- Fast Food
- Family Reunions
I select the ozone layer card because that little strumpet has been trying to get everyone to pay attention to it since the early 90s and I’m sick of it. I place my card face-down so that the judge doesn’t know which card is mine—we want at least to feign impartiality, right? Everyone else does the same. We mix up the cards and the judge chooses which card is most flirtatious. Someone else had the KKK. The judge, because he’s nuts, chooses that one. That was DJ’s card, so he gets the point for the round. Now someone else takes her turn as judge and we continue like that until things get out of hand and we stop. The player with the greatest number of green apple cards at the end wins.
It’s really neat to see the sort of patterns that develop after a couple of rounds have passed. People learn to “play to the judge.” We like to have the judge talk through his choice so that we can get a sense of the convoluted thought processes our friends have. Each round is training. People pick up what works with whom and why others don’t. Strategies emerge. Battles ensue. Complex word associations form. You get the idea.
The way people adapt after a few rounds of the game is basically the same way an artificial neural network (NN) learns. It has some training data. In this case, the training data are the red cards. The problem is to match the right red card with a specified green card. In general an expert will figure out what the right pairings are. The NN will come up with a guess, which is a lot like picked a red card. Then the expert will tell the neural net whether it got the answer right or wrong—in this case, whether it selected the right card or not. In our game, the judge tells us what went wrong during the talk aloud portion of judging. Maybe he thought that only living things could be flirtatious. Perhaps the more bizarre, the better in the judge’s mind. Whatever the case, talking through the solution gives us an opportunity to revise our plan of attack. Next time, we’ll weight one sort of connection over another.
Neural nets do the same thing. They look at the expert’s answer and its answer and compare the difference to calculate the error. The nets use the error to readjust the internal weights so that chance of getting the right answer is better next time. When a net is good at getting a proper response on the training set, then it’s ready to tackle new data that doesn’t yet have a right answer. Apples to Apples is a game that is built around the same premises used to train a neural net. And it’s remarkably fun. But how do we get beyond training; when do we get to have a shot at untamed data? Well, we came up with some new rules.
In this version of the game, the set-up is the same. One person judges against a green card, the rest offer up a matching red card. This time, though, the judge has to guess whose card is which. After playing with the normal rules, you should have a feeling for what sort of answer each person is likely to give. Now you have to work backwards. If you can successfully match a card with a person you get that point. But should you guess wrong, the person who successfully dodged identification gets a point. So, you stand to get a lot of points if you’re the judge. If you’re not, you still could earn one more. And over time, that’s the way to go.
So how did my sister embarrass herself? Well, I was the judge and the word was ‘dirty.’ I said, “Maybe this says a little too much about me, but I’m going to pick ‘handcuffs.’ Whose is it?”
It was my sister’s card. Seeing that, my dad asked, “Janice?”
She floundered a bit. “Well, I like,” she started. The room immediately fell silent. She had the floor. Realizing what she had begun to say, she grasped for words. “At school we were talking, and…”
I cut her off quickly. “Can somebody pass me a cookie?” I asked loudly to DJ, who broke into laughter.
Who knew that computer science could bring a family together like that?
Technorati Tags:apples to apples, family, classification, data mining, neural network, neural net, judging, card game, association, classification, word game, card game, education, learning, training, expert, ozone layer