It’s no secret that machine intelligence has beaten humans in an array of tasks from playing games to face recognition. Despite many advances, AI systems still cannot handle many multiple tasks. For AI experts, the next big aim is to develop a system that can multitask rather than be an expert at only one job.
At Facebook Research, Enter Arthur Szlam and colleagues have started work on an AI assistant skillful of interaction with humans and then executing a wide variety of tasks on request. Significantly, their assistant can learn from its communications and thus increase the range of functions it can perform.
Szlam and associated have chosen Minecraft to chase this goal. Minecraft is a 3D sandbox videogame which enables players to build, craft, explore, and even fight in an exceptional and infinite online world. It is a record-breaking video game of all time, with over 170 million copies sold. More than 90 million people play the game every month.
Despite that Minecraft world offers infinite variety, its simple rules and predictability within specific limits make it useful for AI research. AI researchers have now started to use it to test and train various types of AI systems. Currently, Szlam and team are taking a step further. Rather than a superhuman accomplishment on a sole difficult task, the researchers are interested in proficiency across a wide array of simpler tasks, specified by humans.
According to the team, Minecraft provides a perfect environment for an AI to acquire knowledge from humans. The players enjoy interactions with the assistants as they are developed, generating a rich resource for human-in-the-loop research. Seemingly, the number of actions that are likely in Minecraft is vast. A player could possibly do a broad set of things in a game. If put in simple words, it is all probable ways of placing all the possible blocks into as big a world as fits in RAM, told Szlam and his team.
However, a massive majority of block arrangements in Minecraft are highly improbable. The researchers assume that the allocation of player requests of an assistant will be focused on a small fraction of what is actually possible in the game. This makes the work for an AI quite more uncomplicated. The Facebook team believes that even if their AI system flops at many functions, it will still be valuable. For them, it is not important for the assistant to do well at every possible request. But understanding ordinary language is difficult and challenging.
The Facebook team visualizes the following request:
PLAYER: build a tower 15 blocks tall and then put a giant smiley on top
This is a simple request which requires a significant degree of knowledge. The assistant must comprehend the idea of a “tower” and how to construct one. It should also know that “15 blocks high” is a calculation of the tower height. It also needs to know the term “smiley” and how to make it and identify what “on top” means. Subsequently, these requests could get significantly complicated.
YOU MAY LIKE: Children Getting Affected By AI – What AI Means For Kids
But the chances for an AI to learn are enormous. If Facebook can stimulate enough human help, the assistant could ask humans to help it acquire knowledge.
One of the major criticism of AI systems is that they do not learn new things except those which they are taught and for which data is gathered and prepared. But now Facebook is taking up the assignment of building the AI to self-improve and Minecraft setting is a perfect environment to develop a similar type of learning.
Still at an embryonic stage, the team describes this project as a “call to arms.” It sets up a challenging goal for Facebook. Nonetheless, an initial version of the assistant is available for trial for ordinary users.
They’ve already done a basic version! Its called CraftAssist: https://github.com/facebookresearch/craftassist
Do tell us if you try it out 🙂