Machines at the MALL Learn the Ways of the World
IISc Lab is working to give machines a worldview
Bengaluru: In one corner of the sprawling Indian Institute of Science (IISc) campus lies the seemingly quiet Machine and Language Learning (MALL) Lab.
It’s making a noise in the world of technology, with its grand mission of giving a “worldview” to machines that will enable them to make wider connections between various data and draw the necessary inferences to make decisions. Make no mistake, this is not a random academic pursuit. It is of utmost importance in an era where automation is the next big thing. Automating decisions could save companies huge amounts of money and reduce man power, which can be instead used for more critical tasks.
Recognising this, global technology giants such as IBM and others are funding and collaborating with the lab.
“We, as humans, aggregate a lot of knowledge about an entity and then use the knowledge to make decisions. Increasingly, we are asking machines to make non-trivial decisions on our behalf. But machines do not have the worldview that you and I have,” said Partha Talukdar who is leading the team that works on this project.
“The thesis is that, if you can make that kind of knowledge available to these decision-making agents — to chatbots, robots, personal assistants — then their performance is going to significantly increase,” said the 35-year-old professor.
The team consists of PhDs, master’s students, research assistants and interns, who work on various aspects of this huge task.
The work done at the lab could potentially help in creating knowledgeable personal assistants who can carry out tasks with increased efficiency, come up with algorithms to fulfil taskspecific needs and build programs that can scrounge through open forums to track feedback about products, among others. “IBM Research is looking at task directed knowledge graph,” said Talukdar. The team uses core artificial intelligence (AI) areas such as natural language processing, machine learning and large-scale data analysis methods to build ‘knowledge graphs’. Knowledge graph, a term that Google made famous in 2012, basically tries to understand and learn facts about places, people and things and how these entities are all related. “You can think of that as a graph or a network where the nodes are of entities of interest,” said Talukdar.
Sharmistha Jat, a PhD student who does relation extraction at the lab, gave a peep into what they are doing. “There are different relations, phrases like ‘located in’, ‘prime minister of ’, etc. We want to understand how entities are connected. I look at patterns. For example, if there is a statement saying Narendra Modi was born in Gujarat. The machine needs to know that Gujarat is in India and so Narendra Modi was born in India,” she said.
The knowledge to ‘teach’ machines is gathered mostly through the internet. The internet, with its many social networking sites, blogs and websites, is a repository of information. Though there are certainly labs working on various subsets of the same problem, the ‘comprehensive outlook’ is what makes MALL special, Talukdar said.
“The area that we are working in is unique in the Indian context. There are different pieces of it that are studied in other groups. But the collective view is unique. I do not know any other lab in India and a few globally, that has this comprehensive view,” he said.
Though these are still early days for the lab that began in late 2014, it has already seen varying degrees of success with its collaborations. “There are prototypes that have been developed. Some of the algorithms that we have developed have been implemented on our partner’s data. They have seen different degrees of success,” said Talukdar.
Automating decisions could save cos huge amounts of money and reduce man power, wh ich can be instead used for more critical tasks