“On true general intelligence there has been a lot of work, but the whole scientific community agrees that it will take a very long time before we reach it. It will take decades,” said Victor Muntés-Mulero, the vice-president of Strategic Research, on the sidelines of the company’s Built to Change Summit in the US.
Muntés-Mulero, who is part of the Advance Deep Dive Research initiative of CA, said there was “still a long path to get there”, and in between there had to be lots of work done on various aspects of intelligence.
Maria Valez-Rojas, a research scientist with the Deep Dive Research, said the safety issues coming up in robotics, and the need to solve all of them, meant it wasn’t going to happen in the near future.
Many leading commentators have mentioned from time to time that human intelligence in machines is less than a decade away.
Scientists now realise that making machines is easy, but bringing in the whole gamut of ethics, human rights and social construct will be difficult to imbed.
In this context, Muntés-Mulero said they were researching how to make AI behave ethically, adding that some intelligent software recently turned out to be wrong.
For example, Google Translate often became sexist when translating sentences, say from Turkish into English, stating “he’s an engineer or she’s a nurse”, when the original did not imply a gender. Similarly, AI delivery systems often discriminated against residences of people of colour.
“You would like AI to do what you intend it to do, but sometimes it can behave unexpectedly,” he said, adding that Alexa or Siri could take non-human instructions from audio in television and deliver results which were unintended.
“Your Apple watch can be hacked to get your passwords by motion detection as you type them into the laptop,” he said.
Human speech and sense of fairness and human rights issues were important, but it would take a while before machines were taught to discern them. In that context, he quoted Facebook chief executive Mark Zuckerberg, who said that it was easier for AI to discern the image of a nipple than to detect hate speech.
Muntés-Mulero said things became more complicated when you realised that there was no single definition of fairness or social rules and that they sometimes differed from country to country.
Defining all that mathematically was a challenging task in itself, he said. But it was their goal, in research, to create a better world by ensuring fairness and appreciation of human rights in AI.
Valez-Rojas, who’s working on “cobotics”, or human collaboration with robotics, said the new era of interaction with robots was coming where they would collaborate “not to replace us, but to help with our jobs”.
She said AI would eventually help robots to understand the human environment. The interaction could lead to problems because “humans are so unpredictable”.
Also, it would take a lot of effort to teach robots how to differentiate between concepts.
“When you tell a robot to ‘clean up the mess’, it might proceed beyond the coffee spilt on the floor and wipe out the massive calculations that you have done the whole week on the blackboard. Because that may mean ‘mess’ for it.”
She said humans had moved on from robots which did repetitive work to controlled robots like drones to somewhat intelligent ones like those that could walk across the room and avoid obstacles on their own.
“Finally, we have to make robots that can behave autonomously and help humans in their jobs,” she said.
Steven Greenspan, a research scientist, said a lot of research was being done at CA on security and privacy which were the top concerns in the world.
He said General Data Protection Regulation (GDPR), passed in Europe, had far-reaching consequences for not only companies there, but almost any company in the world that did business with Europe.
CA was working on products where the requirements were embedded into them so companies did not have to be experts in understanding what GDPR required.
The thrust of the new regulations provided that the individual was the owner of the data and could decide who it should be shared with, and could take it back whenever they so desired.
Greenspan said access control had its problem since passwords could be cracked, hacked or stolen. In that context, they were researching “behavioural markers” which would better identify the person using any software.
“We can figure out how you type or what your pattern of walk is with your mobile, even if you are erratic in your behaviour. Even if you go out for lunch or sometimes skip lunch to work through, it’s all unique.
“If you step out after logging in, and someone else comes and uses the system, we can know.”
He added that by using the behaviour markers, they had been able to get correct identification of more than 95%.
The research is also looking into whether they identify where intent changes in a person, from benign to malicious. - IANS