Header Ads

New AI Model to Aid Future Robots Fit in our Settings

In a development that has kindled hopes of making robots that are more acceptable in the future, scientists have a developed a new Artificial Intelligence (AI) model that can play ‘Pictionary,’ a social game. This is the first computer-based model for this game.

The team comprises Ravi Kiran Sarvadevabhatla, an assistant professor at IIIT-H, affiliated with Kohli Center on Intelligent Systems, Trisha Mittal, an MS-Computer Science student at University of Maryland, and Shiv Surya, an applied scientist at e-commerce firm Amazon, all former members of Prof Venkatesh Babu’s Video Analytics Lab in IISc want models that can perform in such games to become the benchmark in AI.

So far, computer-based modelling of human player games such as Backgammon and Chess has been an important research area and accomplishments of game engines— TD-Gammon, DeepBlue, AlphaGo, for example—and their ability to mimic human-like game moves has been a well-accepted proxy for gauging progress in AI.

But all these are zero-sum games where one wins and another loses, they are popular and rules for a win are well-defined.

“...AI in such games are always designed to win, Pictionary is not like that. In Pictionary, the notion of win is blurry as the judge can accept multiple interpretations of what is being drawn, making it very abstract,” Sarvadevabhatla told TOI.

In social games, the primary emphasis is on co-operative game-play in a relaxed setting, unlike the typical zero-sum games. For instance, if a person is drawing badly, the correct thing for the robot to do is to guess wrong as humans would do, Sarvadevabhatla argued, adding that the robot must thus be taught to do this, which is challenging.

It requires a rethink in terms of agent and gameplay modelling, and Pictionary, in particular, is complicated by additional factors such as multimodal gameplay—guesser uses speech/lexical modality while drawer uses visual modality—asynchronous turn-taking and high-level notions of what constitutes a ‘win’.

To address these challenges, the team designed a deep network-based approach as a recurrent neural network which maps the temporally and incrementally evolving sketch to a sequence of human-like object guesses.

Developing such computational models of human guessing enables characterization of realistic, possibly suboptimal human drawings and responses which arise in Pictionary.

“The performance of these models measured relative to human responses enables us to quantify the extent to which these models encode non-trivial human behaviour and mimic human responses. More broadly, our work represents a preliminary but important step towards the realization of AI agents whose human-like, relatable responses increase their acceptance in social settings," the team said.

Pointing out that a robot/model’s performance on multi-disciplinary tasks such as Visual Question Answering (VQA) is considered a marker for gauging progress in Computer Vision, the team said its model also has the first Sketch-QA. Sketch-QA involves asking a fixed question like “What object is being drawn?” and gathering open-ended guess-words from human guessers.

The resulting dataset was then analysed and to mimic Pictionary-style guessing, the team proposed the deep neural model which generates guess-words in response to temporally evolving human-drawn sketches.

“Our model even makes human-like mistakes while guessing, thus amplifying the human mimicry factor. And, in the future, we will have models that can even draw,” Sarvadevbhatla said, adding that models can now be developed for other similar social games too.
Powered by Blogger.