I always wondered if ‘growing’ a robot was the best method to archive AI. Wired examine a form of this in The Power of Babble
MIT researcher Deb Roy is videotaping every waking minute of his infant son’s first 3 years of life. His ultimate goal: teach a robot to talk.
The idea was to supplement his robot’s long-term memory with short-term memory. Both would be engaged in pattern recognition, searching speech input for recurring phonemes, but the short-term memory would focus on the recent past . By giving Toco a mild case of ADD, Roy made his robot more like the kids he was trying to emulate. Without the ability to prioritize recent experience, Toco’s search algorithm had been spending valuable time cycling through every phoneme it had ever encountered.
And with the addition of short-term focus? Roy found that Toco could learn much faster if it were allowed to concentrate on the ball or the cup. Taking input directly from the baby lab — raw audio that the machine “hears” by analyzing the sound’s spectrograph — Toco was building an elementary vocabulary. “It caused quite a stir,” Roy says. “This was the first time that a computer took a lot of audio input without a lot of massaging.”