The director of the Artificial Intelligence Research Institute at the CSIC, Ramón López de Mántaras, talks about the progress of automated learning and points out that common sense is the biggest challenge for AI.
“It is said that within 30 years machines will be smarter than us. That they will be superior to us. In reality, we are currently developing specific intelligences. The system performs a task very well, better even than human beings: there are systems that are able to make a better diagnosis than a doctor but they do not possess general medical knowledge; or they can beat us at a specific game but are unable to play any other game. Artificial intelligence is not around the corner.”
Ramón López de Mántaras, the director of the Artificial Intelligence Research Institute at the CSIC, has no doubts about the potential of AI. From his office at Universidad Autónoma de Barcelona, he qualifies the progress of the branch of computer science that develops programs to emulate human beings in terms of their learning and logical thinking.
According to López de Mántaras, common sense is the biggest challenge faced by artificial intelligence. “The general ‘question-answer’ interaction with deep semantic exchange is not impossible but I think it will take a while; and I mean decades. There are specific algorithms that can do plenty of different things but they are independent from each other.”
The researcher stresses that in a specific area (e.g. banking) it is possible to hold an intelligent conversation. “If 90% of people are going to, for example, use bank-related terms online, the machine is able to represent and anticipate the concepts that will be used, establish relationships and add semantic depth. From there, the robot can hold what you could call an intelligent conversation. However, if the terms change, the machine is unable to respond. When you move out of a specific framework, the conversation is over.”
The problem arises when general concepts are involved: “Two examples: Google Translate is pure statistics and, for this reason, it’s a pretty stupid and fragile system. Some of its translations are a disaster. Also, there is no reasoning behind artificial vision and, as a consequence, when we ask the machine to interpret the image of a baby with a tooth brush in front of its face, it tells us that the baby is holding a baseball bat. There are millions of photos of children with baseball bats online. It’s a photo you can find anywhere and that’s what the system learns. This is the weakness of a system based on mass data analysis.”
And where is artificial intelligence less fragile? The researcher mentions two areas – reinforcement learning, e.g. logistics and its assembly lines; and inductive learning, e.g. the robot recognizes chairs: even if it is seeing chairs that it has never seen before, it identifies common characteristics since it is able to generalize.
In terms of vision and language, artificial intelligence is progressing less because of the high number of examples to be learned. Perception and communication are very complicated since machines possess no implicit knowledge or experiences: “Learning based on mass data analysis works when you use millions of examples – a system recognizes a cat because it has seen millions of cats; a child only needs to see a cat once to know it is a cat,” added López de Mántaras.
In terms of vision and language, artificial intelligence is progressing less because of the high number of examples to be learned
Ethics of artificial intelligence
The physicist is very hard on the aura surrounding AI: “When we talk about artificial intelligence, we are talking about very specific intelligences. It’s not what you see in the movies or what you hear about: people who talk about post-humanism as a consequence of technological singularity don’t really know what they are talking about since none of the people selling the idea that we will become obsolete are experts in artificial intelligence. I don’t agree at all that artificial intelligence means the end of Humanity.”
Ethics is also a concern for the researcher. He points out that it is important to have regulations and make brave decisions. “We should ban bots that buy and sell in the stock markets in milliseconds or less because they take over the system and human traders don’t have anything to do.”
“There are already competing software companies and it is unacceptable that they dominate the lines of communication and make it impossible for other companies to enter the market. And they also push human buyers out. They are very dangerous and destabilize the economy,” he warned.
While he waits for regulations, his team of 60 at CSIC keeps working on a single system that integrates all capacities of intelligence so that robots can perform increasingly more complex tasks.