Artificial intelligence is getting smarter
Computer scientists have made significant progress in the development of AI in the past few years
ARTIFICIAL intelligence (AI) plays a major role in the fourth industrial revolution (4IR) and will increasingly change our future in the years to come.
A mere four years ago AI was not even able to pass a Grade eight science test. Seven hundred computer scientists competed in a contest with a significant amount of money as prize. They had to build artificial intelligence that could pass a Grade eight science test.
The computer scientists did their best, but not even the most advanced AI system could score better than 60 percent in the test. It seems that the AI was just not advanced enough to fully reach the language and logic skills expected of students in the eighth grade. But this all changed in early September of this year when a renowned laboratory in Seattle, US, – the Allen Institute for Artificial Intelligence – demonstrated a new AI system that has been able to pass the Grade eight exam with great ease.
The AI system was able to answer more than 90 percent of the questions correctly. And if this was not enough, it also passed the Grade 12 science exam with an amazing 83 percent!
The AI system, named Aristo by researchers, is a dedicated system specifically built for multiple-choice type of questions. Although it took the standard exams for students of New York, all questions with pictures or diagrams, as well as direct answer questions were removed, since that would also require computer vision skills in combination with language understanding and logic skills.
Although the questions were multiple-choice, only some of them required mere information retrieval, while the rest required logical reasoning by the AI system.
The Allen Institute, founded by the Microsoft co-founder Paul Allen, started with their AI research already in 2013. Instead of the typical AI benchmark tests like chess or other games, the researchers at the Allen Institute decided to rather test the AI on standardised science tests, because a science test cannot be mastered merely by learning a set of fixed rules.
The AI must also be able to do logical reasoning to make certain connections and the right choice.
Neural networks – complex mathematical systems that can learn tasks by analysing enormous amounts of data – are the main drivers of the work of the Allen Institute of Artificial Intelligence. Through the recognition of patterns in thousands of objects, a neural network can learn to recognise the various objects with great accuracy.
Similarly, AI systems can learn the finer nuances of a language through analysing thousands of articles and books that were written by humans.
It is apparent from this breakthrough that computer scientists have made significant progress in the development of AI in the past few years, in particular with regard to AI that can understand languages and imitate the logic and decision-making processes of human beings.
Many of the world’s leading research laboratories have made impressive progress with regard to the ability of AI machines to understand and respond to natural language. AI machines are continuously improving in the finding of information, analysing of documents, answering of questions and even the generating of language on their own.
It is never easy to complete another person’s thought since it requires some advance reasoning capabilities and an excellent command of language. But the Allen Institute for Artificial Intelligence were successful to build AI systems that could begin to understand the intricacies of natural language and that could thus independently complete sentences in English.
The original work was done at Google, where researchers built a system called Bert, that analysed thousands of Wikipedia articles and a large library of romance novels, science fiction and other books. Through the analysing of all this material and learning about the fundamental ways language is constructed, Bert was able to learn how to guess the missing word in a sentence.
The Aristo system of the Allen Institute was built on top of the Bert technology. After feeding Bert a huge number of questions and answers, it learned to answer similar questions on its own.
Significant breakthroughs have been made with regard to natural language processing. Many of the newer AI systems using language models are used in research projects. For instance, one of these projects entails conversational systems and tools that are designed to identify fake news that are so prevalent today.
Another project involves AI assistants such as the well known Alexa that plays music at home and looks up certain information; Siri that controls many functions on the iPhone; or Bixby on the Samsung that could provide weather and other context relevant information. Unfortunately, these intelligent assistants have not fully lived up to all the hype around them.
The promise was that they would simplify our lives, but it did not really happen, since they could only recognise a very narrow and pre-defined range of directions.
But new techniques that capture semantic relationships between words and enable machines to better understand natural language are about to expand digital assistants’ repertoire.
Researchers at OpenAI already in 2018 developed a technique that trains an AI system on unlabelled text to avoid the expense and time of categorising and tagging all the data manually. These improvements, together with advances in speech synthesis, allow us to move from giving AI assistants limited and simple commands to having reasonable conversations.
Soon the AI assistants will be able to deal with a variety of issues every day, such as taking minutes of a meeting, finding information relevant to the situation, making or changing appointments.
Google Duplex, the somewhat human-like upgrade of Google Assistant, can even pick up your calls to screen for spammers and telemarketers. It is also able to make calls on your behalf, schedule restaurant reservations and make other appointments.
In China the well-known Chinese multinational conglomerate company Alibaba, specialising in e-commerce and retail, are using a smart chatbot called AliMe, which co-ordinates the delivery of orders over the phone and even haggles about the price of goods over chat. At the very foundation of AliMe – an intelligent human-computer interaction system – lies natural language processing.
Despite all the advances, some scientists are still very sceptical regarding the progress made by Aristo and other AI systems. Jeremy Howard, the Australian entrepreneur and chief executive of Fast.ai, another influential AI laboratory in San Francisco (US), is of the opinion we are still many years from completely mastering natural language or duplicating true human intelligence.
According to the sceptics, current AI technology cannot be compared to real human students and their ability in reading comprehension or logical reasoning. All these AI systems are designed for narrow tasks and lack common sense. Language is also only one piece of the puzzle.
Although AI still needs to improve tremendously to be compared to human intelligence, the advances made by Aristo and other AI initiatives are significant and, according to Professor Oren Etzioni, the chief executive of the Allen Institute of Artificial Intelligence, could spread to a range of products and services, such as Internet search engines and hospital record-keeping systems.
With constantly increasing data and computing power AI technology will keep on improving. The AI research on natural language processing could therefore indeed lead to systems that can manage a full conversation so that in the future we could have a proper conversation with our AI machine or robot if we are alone!