Israeli expert calls for machine ‘Wikipedia’ to help AI computers learn
Challenges are seen to abound for the future of artificial intelligence, especially in natural language processing and transparency over where the machine’s data is coming from
Shoshanna Solomon was The Times of Israel's Startups and Business reporter
An Israeli expert in artificial intelligence — the field that gives computers the ability to learn — is calling for the setting up of a Wikipedia-style repository of knowledge for smart machines, as a group of his peers acknowledged that challenges lie ahead for the burgeoning field.
“We are still at the very beginning of AI but AI is already everywhere,” said Aya Soffer, world vice president of AI technologies, IBM. However, its applications are still very “narrow,” she said, with the machines requiring very well-defined tasks and a lot of data in order to accomplish them.
Although the concept of AI has been around since the 1950s, it is now enjoying a resurgence made possible by the higher computational power of chips. The field is expected to grow at a compounded annual rate of almost 37 percent from 2018, and is expected to be a $191 billion market by 2025, according to MarketsandMarkets, a research firm.
Artificial intelligence and machine learning are used for a wide range of applications, from facial recognition to detection of diseases in medical images to global competitions in games such as chess and Go.
Therefore, companies are trying to create a middle-of-the-way AI, called “broad AI,” in which machines can do more tasks using less data, said IBM’s Soffer.
“In the real world, most companies do not have a lot of data,” she said. One of the biggest challenges ahead for AI technology is to have machines that can learn with access to smaller amounts of data. Feeding huge amounts of data to the machines takes time, data availability is limited, and data changes all the time, she said.
One solution, she said, is to try to get AI machines to do “data augmentation” themselves, meaning use AI to create new data using the existing information.
Another solution was proposed by Prof. Moshe BenBassat, an artificial intelligence researcher at IDC and the organizer of the conference, who said that “there needs to be an international effort” by the AI community to create a special database that smart machines will be able to access, understand and use.
BenBassat is the founder of Clicksoftware Technologies Ltd., a firm that applies complex algorithms and AI to help manage workers, and Plataine Ltd., a maker of AI solutions for advanced manufacturing in a variety of industries.
If a smart machine needs knowledge beyond the training it gets directly from humans, it should be able to access expert data as well, he said, just as humans can. “So why not build something like Wikipedia, but structured in a way that smart machines can use and understand?”
BenBassat calls his proposed platform the ReKopedia. “I am talking about the deep knowledge” that is needed for training physicians, for example, he told The Times of Israel on the sidelines of the conference. This is knowledge that AI machines cannot learn just from looking at patient data, but will allow “smart machines to make deeper inferences,” he said.
The idea would be to develop software structures, like neural networks, Bayesian networks and decision trees, that hold the “knowledge of humanity” and that smart machines can understand, he said.
This is a challenge for the AI community, he said. “It will amplify AI capabilities in a substantial manner.”
To trust or not to trust
An additional challenge ahead is gaining users’ trust, said IBM’s Soffer.
Even if the technology gets to be 99% accurate, people are still hesitant to use AI systems because they don’t know whether the decisions reached by the machines are based on fair or biased data, she said.
What’s needed, she said, is more transparency. AI technology should come with a fact sheet, along the lines of a food label. This is something that IBM is starting to do with its technologies: providing an accounting of how the machine has been trained, what data it has been trained on and its level of accuracy.
This will help “establish more fairness and trust,” she said.
The inability to understand human language, with all its nuances, metaphors and contexts, is another reason AI is not yet achieving its potential.
“AI is very far from truly understanding natural language,” she said. Machines can translate very well, Soffer said, but only because they don’t really need to understand anything about the text itself in order to do so.
There is still a big gap between the language capability of AI machines and that of young children, she said.
“If we are not able to combine AI with reasoning and knowledge,” the potential of the technology will remain unfulfilled, she warned.
The AI conference, held on Sunday at the IDC Interdisciplinary Center, Herzliya, a private college, focused on applications of artificial intelligence in cyber, medicine, autonomous cars, design and production, and industry.
About a hundred of Google’s thousand workers in Israel work on AI developments, said Yossi Matias, a VP Engineering and managing director of Google’s R&D Center in Israel. He noted that among the team’s projects is the Call Screen, launched by Google in September for some phones in the US, which uses speech recognition and text-to-speech AI tech to identify a caller when the caller’s number is not recognized.
