Hal, like any 18-month old baby, is learning the rudiments of speech. He talks about red balls and blue balls, knows his Mommy and Daddy, and likes to go to the park. A child development specialist was given transcripts of Hal's conversations with his caretakers and declared him a healthy, normal little boy. What she wasn't told is that Hal is a computer program running on a regular Windows PC.
Ai uses behaviorist principles to teach our child machine - nicknamed Hal - to hold a conversation. Our approach was outlined by the computing theory pioneer Alan Turing in his 1950 article Computing Machinery and Intelligence. Turing viewed language as the defining element of intelligence; he believed that by giving a machine the capacity to learn, and a willingness to ask questions, you could "raise" an intelligence, an entity capable of rational, engaging conversation.
Education
Learn how Hal is being educated, and read about his training process.
State of Mind
The Ai child machine is built on a statistical model of language, coupled with advanced learning algorithms.
First Words
HAL is developed to meet basic human language development milestones. Look inside to find out what HAL's talking about.
Tuesday, March 8, 2011
Research plan (Artifical Intelligence)
Ai has developed a research program aimed at true artificial intelligence - allowing people to converse with their computers in everyday language.
In teaching a computer to use language, Ai takes a scheduled, developmental approach, applying the behaviorist model of learning.
Our research plan is based on an iterative cycle, designed to improve the language skills of the system with each software update ("brain upgrade"). The developmental milestones we set for our child machine are based on human language-use milestones, with progress being evaluated by experts in child development
Applying the principles of behaviorism, we teach language to the child-machine through a system of rewards and punishments. The child-machine thus learns to use language, rather than having language built into it. Subjective Intelligence Related Article
At the Ai research facility, trainers converse with the machine, engaging it in conversation and monitoring its progress. The trainers test the limits of the child-machine's intelligence, and share their assessments with the algorithm developers. The developers consequently update and adapt the child-machine's algorithm, or "brain", making it more "human" in its language capability. Every so often, a new version of the brain is handed to the trainers, and the process repeats.
As the trainer works with the child machine, he or she frequently reports back to the developers on its progress. Our metric for success is clearly defined as "the language capability of a human of a defined age." For the first several iterations of the child machine, we sought to have the child machine speak at the same developmental level as a 15-month old. Now we are working on raising the child to 18 months. The actual time spent training the child does not correspond to its age; rather, these measurements of linguistic ability are standard, accepted guidelines for determining if a child is making linguistic progress or not.
In teaching a computer to use language, Ai takes a scheduled, developmental approach, applying the behaviorist model of learning.
Our research plan is based on an iterative cycle, designed to improve the language skills of the system with each software update ("brain upgrade"). The developmental milestones we set for our child machine are based on human language-use milestones, with progress being evaluated by experts in child development
Applying the principles of behaviorism, we teach language to the child-machine through a system of rewards and punishments. The child-machine thus learns to use language, rather than having language built into it. Subjective Intelligence Related Article
At the Ai research facility, trainers converse with the machine, engaging it in conversation and monitoring its progress. The trainers test the limits of the child-machine's intelligence, and share their assessments with the algorithm developers. The developers consequently update and adapt the child-machine's algorithm, or "brain", making it more "human" in its language capability. Every so often, a new version of the brain is handed to the trainers, and the process repeats.
As the trainer works with the child machine, he or she frequently reports back to the developers on its progress. Our metric for success is clearly defined as "the language capability of a human of a defined age." For the first several iterations of the child machine, we sought to have the child machine speak at the same developmental level as a 15-month old. Now we are working on raising the child to 18 months. The actual time spent training the child does not correspond to its age; rather, these measurements of linguistic ability are standard, accepted guidelines for determining if a child is making linguistic progress or not.
Subscribe to:
Posts (Atom)