When people think of artificial intelligence, chances are the images they conjure up are not pleasant. Whether it’s HAL9000’s murderous insolence in 2001: A Space Odyssey, existentialist questions about the nature of life in Ex Machina or the full-fledged genocide of humanity at the claws of killer robots in the Terminator franchise, people have been taught for decades to fear and distrust artificial intelligence.
What they probably don’t realize is that they more than likely use this technology every day.
Artificial neural networks, a rudimentary form of artificial intelligence, are present in practically every search engine, translator, and virtual assistant on the planet. Type something into Google? Neural networks search through billions of webpages in seconds to find the most relevant result. Every time Siri, Cortana, or the innovatively-named Google Assistant are given a request, neural network technology goes to work.
Neural networks most recently came into the public eye in February’s ‘deep fakes’ controversy, where neural network technology was used to superimpose the faces of celebrities onto pornographic material.
How do they work?
Artificial neural networks are computer systems that simulate the human brain’s capacity to learn and adapt—a process known as deep learning.
Neural networks are fed information and taught to perform certain abstract tasks, before then applying that knowledge to the intended problem. This allows them to tackle problems too complex for a human mind and too versatile for a traditional computer system.
Inspired by Canadian neuropsychologist Donald O. Hebb’s model of the human brain’s learning abilities in the 1940s, neural networks were tested in computing systems during the 1950s and 1960s. They were steadily developed until the 1980s, when interest in neural networks came to an abrupt halt, as they were largely discredited by the scientific community of the time. Funding went elsewhere, and experts concerned themselves with the tried and true methods of their respective fields.
In the late 1990s and early 2000s, new developments in computer processing power meant that artificial neural networks were finally a significantly viable idea. It was during this time that the practical applications of neural networks became clear.
Trailblazers such as Qi-Jun Zhang, associate dean (research) in the Faculty of Engineering and Design at Carleton, jumped on this opportunity, adapting neural network technology into their respective fields.
Neural Networks for RF and Microwave Design, authored by Zhang and the late K.C. Gupta of the University of Colorado, boasts being the “first book of its kind” to educate readers on how neural networks can be used in designing high-frequency electronics. This is a field with many complex relationships which cannot easily be simulated by conventional computer processing.
“A lot of behaviours in high-frequency electronics cannot be described by conventional physics and engineering laws. They were better at low frequencies, but at higher frequencies, it’s not that easy. That’s where we use neural networks. We combine knowledge with neural network learning capabilities, leading to what we call knowledge-based neural networks,” Zhang said.
Although Zhang and other like-minded innovators found a use for the new technology, many in his field did not. According to Zhang, five to 10 years ago, the engineering community was still largely skeptical about its usefulness.
“People didn’t expect neural networks and learning machines; they expected solid engineering with results,” he said.
This was true of many computing fields. In the field of computational linguistics, also known as natural language processing, neural networks were dismissed as a passing fad.
“It was only a few years ago that everyone was laughing about these kinds of services, but now people use them very frequently in everyday life, and they’re widely accepted,” Jörg Tiedemann, a professor in the University of Helsinki’s department of linguistics and philology, said.
Using Neural networks
Tiedemann’s specialty is machine translation, which is concerned with translating speech and text between languages. Research in this field is regularly applied to translation and language localization technology.
“We get a bit more attention from the surrounding society now. I think that’s made it easier for us,” Tiedemann said. “Now everyone knows what language technology is about—up to a few years ago, you still had to explain what is the use for all of this, whereas now you can always point to some existing tools, which everyone knows, and you can explain that this is what we are working on.”
With his research, Tiedemann said he hopes to someday see a world without language barriers.
“What I see for the future is that no one will think twice about accessing multilingual content. Right now, we have a sort of discrimination based on linguistic barriers. You want to give all people in the world the same possibilities,” he said. “There will always be a lot of differences with new technology, but I hope that by taking away language barriers and adding communication possibilities that you can access somewhere else, you will balance the possibilities more evenly.”
Tiedemann’s vision is not far off, with Microsoft’s Skype featuring a real-time translation application for over 50 languages, and Google’s voice and image recognition service boasting a 95 per cent accuracy rate.
Sabine Nyholm, a computational linguistics graduate student working under Tiedemann, said neural network technology could be adapted for language localization in movies, television, and video games.
“Localisation is lagging behind a lot, and many of the tools they use are archaic compared to what would be possible or available with modern language tech,” Nyholm said. “There’s a load of untapped potential there.”
Besides engineering and language technology, neural networks are used in practically all fields that require information processing.
The November 2016 issue of Applied Sciences, a monthly science journal, details how neural networks could be used in fields such as chemistry, biology, ecology, neurology, and physics.
“It’s definitely a subject of the future. It can be used for anything which has some kind of data available, which is essentially everything. It should make a lot of jobs easier,” Nyholm said.
From manual to automated
One of the main concerns with the development of machine learning technology is its potential to replace human workers. Zhang, however, said there is nothing to worry about.
“It’s more about enabling humans. Neural networks give people more powerful help,” he said.
According to Zhang, neural networks are the next step in creating tools for humans to use. This technology can be applied to create better models and more efficient simulation systems.
Although the neural networks applied in the field at the moment are advanced multilevel networks, anyone can build a simple neural network with a little help.
Neural Network Programming with Java, authored by Fábio M. Soares and Alan M.F. Souza, aims to teach people with little or no programming experience to build neural networks with Java, a common programming language.
“Many people use simulators to simulate engineering design. There’s not enough intelligence to them right now. Our research goal is to add intelligence to those tools so that they provide a more intelligent help,” Zhang said.
Tiedemann, on the other hand, said he believes deep learning technology holds the potential to make certain jobs obsolete, but considers it to be an issue of human adaptation.
“It’s exactly the same as it’s always been. Technology develops and people are always afraid that ‘this will take away my job.’ It does, of course. It takes away jobs that you cover with technology. But as we’ve seen in history, there’s always new kinds of jobs coming up that have different qualifications and needs,” he said.
A question of ethics
As neural networks have become increasingly common, ethical questions about their usage are being raised.
“We are now able to process a lot of data automatically, and there are algorithms potentially making some decisions, [and] drawing some conclusions. You can imagine how easily that can be misused for different purposes. If you leave decisions to a machine, and you just trust that, it can have some negative impacts,” Tiedemann said.
He said his concerns revolve around the potential of neural networks to make decisions which influence people, especially if there is a corporate interest involved.
“Information flow is actually quite regulated by the algorithms that run in these commercial providers—they decide what kind of information you get, and that really influences our behaviour. That’s quite the risk. Looking at the American elections, how easy it is to influence people by just putting certain opinions into the data flow,” he said.
Tiedemann advised that neural networks must be developed openly and transparently, to prevent misuse. Otherwise, development could be pushed in a direction with the potential for harm.
“It’s really important that we have open development and open science so that we don’t leave it to the companies to develop those techniques,” Tiedemann said. “If there’s a commercial interest that drives the development and there’s no transparency then we are trapped, because we rely on their services without knowing how this influences us.”
We would do well to heed this warning.
Rather than the oppressive robot overlords seen in film, the rudimentary artificial intelligence which has arrived is a nurturing force. Although neural networks have great potential for human advancement, they also have a great potential for harm, albeit, not the expected physical kind. Neural networks could elevate a field to never before reached heights. Alternatively, they could be the most efficient means of corporate exploitation ever seen.
As Tiedemann said of neural network based technology, it’s not so much in the future any more.
Whatever the outcome, neural networks are here, and they’re here to stay. What happens next is in our hands.
Graphic by Manoj Thayalan
