Facebook robots knew too much


Facebook has shut down a controversial chatbot experiment that saw two AIs develop their own language to communicate.
The social media firm was experimenting with teaching two chatbots, Alice and Bob, how to negotiate with one another.
However, researchers at the Facebook Artificial Intelligence Researchers (Fair) found that they had deviated from script and were inventing new phrases without any human input. The bots were attempting to imitate human speech when they developed their own machine language spontaneously at which point Facebook decided to shut them down.
“Our interest was having bots which could talk to people,” Mike Lewis of Facebook’s Fair programme said.
The researchers were teaching the chatbots, artificial intelligence programs that carry out automated one to one tasks, to make deals with one another.
As part of the learning process they set up two bots, known as dialogue agents, to teach each other about human speech using machine learning algorithms. The bots were originally left alone to develop their conversational skills. When the experimenters returned, they found that the AI software had begun to deviate from normal speech. Instead, they were using a brand new language created without any input from their human supervisors.
The new language was more efficient for communication between the bots, but was not helpful in achieving the task they had been set.
“Agents will drift off understandable language and invent code words for themselves,” Dhruv Batra, a visiting research scientist from Georgia Tech at Fair said.
“Like if I say ‘the’ five times, you interpret that to mean I want five copies of this item. This isn’t so different from the way communities of humans create shorthand.”
The programmers had to alter the way the machines learned language to complete their negotiation training. “During reinforcement learning, the agent attempts to improve its parameters from conversations with another agent,” a spokesman said. “While the other agent could be a human, Fair used a fixed supervised model that was trained to imitate humans.”
The second model is fixed, because the researchers found that updating the parameters of both agents led to divergence from human language as the agents developed their own language for negotiating.
“Facebook’s artificial intelligence researchers announced last week they had broken new ground by giving the bots the ability to negotiate, and make compromises. The technology pushes forward the ability to create bots “that can reason, converse and negotiate, all key steps in building a personalised digital assistant,” Lewis Batra said in a blog post.
Up to now, most bots or chatbots have had only the ability to hold short conversations and perform simple tasks like booking a restaurant table. But in the latest code developed by Facebook, bots will be able to dialogue and “to engage in start-to-finish negotiations with other bots or people while arriving at common decisions or outcomes,” they wrote. The Fair team gave bots this ability by estimating the “value” of an item and inferring how much that is worth to each party.
The bots can also find ways to be sneaky. In some cases, bots “initially feigned interest in a valueless item, only to later ‘compromise’ by conceding it an effective negotiating tactic that people use regularly,” the researchers said. This behaviour was not programmed by the researchers “but was discovered by the bot as a method for trying to achieve its goals,” they said. The bots were also trained to never give up. “The new agents held longer conversations with humans, in turn accepting deals less quickly.
“While people can sometimes walk away with no deal, the model in this experiment negotiates until it achieves a successful outcome.” PA

No posts to display