SM/AIUEO/Getty

PONG is a gloriously simple video game: you control one paddle, aiming to bounce the ball past your opponent’s paddle. Artificial intelligence has learned to play it so well that it can easily beat human players. But try to get the same AI to play Breakout, a very similar paddle-based game, and it is utterly stumped. It can’t reuse what it has learned about paddles and balls from Pong, and has to learn to play from scratch.

This problem dogs modern artificial intelligence. Computers can learn without our guidance, but the knowledge they acquire is meaningless beyond the problem they are set. They are like a child who, having learned to drink from a bottle, cannot even begin to imagine how to drink from a cup.

“A computer is like a child who learns to drink from a bottle but cannot imagine how to drink from a cup“

At Imperial College London, Murray Shanahan and colleagues are working on a way around this problem using an old, unfashionable technique called symbolic AI. “Basically this meant an engineer labelled everything for the AI,” says Shanahan. His idea is to combine this with modern machine learning.

Symbolic AI never took off, because manually describing everything quickly proved overwhelming. Modern AI has overcome that problem by using neural networks, which learn their own representations of the world around them. “They decide what is salient,” says Marta Garnelo, also at Imperial College.

Neural networks have delivered the big AI advances of recent times, but the representations they use are incomprehensible to humans and can’t be transferred to other neural nets. So for each fresh task, neural networks must build new ones. They learn slowly, relying on big data to chew on and plenty of processing power.

Shanahan’s work aims to tie symbolic AI to the autonomous learning of neural networks, allowing some knowledge to transfer between tasks. The prize is learning that is quick and requires less data about the world. As Andrej Karpathy, a machine learning researcher with the firm Open AI, put it in a recent blog post: “I don’t have to actually experience crashing my car into a wall a few hundred times before I slowly start avoiding to do so.”

Symbolic AI also helps us understand how machines make decisions, something we often can’t do. “Neural networks don’t convert the reality around them into the kinds of symbols that we use,” says Joanna Bryson, an AI researcher at the University of Bath, UK. By “symbols”, Bryson and other AI researchers mean any kind of reusable concepts or labels, such as words or phrases.

Shanahan and Garnelo’s hybrid architecture retains neural networks’ ability to interpret the world independently. However, the researchers combine that with some basic assumptions that reflect the way we understand the world: things don’t usually wink out of existence for no reason; objects tend to have certain attributes like colour and shape. This allows the hybrid to build rudimentary common sense. “Our little system very quickly learns a set of rules,” says Shanahan. These let it handle unseen situations that are beyond a purely neural-network-based system.

The team tested the hybrid’s abilities on a simple board game. A mix between tic-tac-toe and Pacman, it features a cursor moving around a board littered with noughts and crosses. Hitting a 0 or × scores or loses a point respectively. Crucially, the distribution of the symbols is different every time, and the hybrid AI had to work out what actions were associated with reward. “If I go get that 0, that’s good. If I go get that ×, it’s bad,” says Shanahan.

When pitted against “Deep Q-Network” (DQN), an algorithm created by Google’s subsidiary DeepMind, the AI did extremely well, beating its score on randomly generated boards that neither architecture had seen before (arxiv.org/abs/1609.05518).

Crucially, the hybrid was able to transfer what it had learned across games. After 1000 training sessions, DQN managed a positive score on half of its games. But it took the hybrid only 200 sessions to arrive at a strategy that earned a positive score on 70 per cent of its games. Shanahan puts it down to it being able to port a rudimentary strategy across different games.

“I don’t want to hype this up too much,” says Shanahan. “It is just a prototype. The game is simple, and the hybrid beat an old version of DQN.”

Still, the implications of transferable learning are fairly significant. “Being able to pick up regularities at different levels is an important component of human-like intelligence,” says Bryson.

This kind of hybrid learning is important for robotics. Powerful learning that involves many layers of neural networks is hard to apply there because of the volume of data needed, says Coline Devin, a computer scientist at the University of California, Berkeley.

Devin sees hybrid architectures as having a particular advantage for driverless cars. “They could use deep learning to process camera images,” she says, while accessing a library of preset rules – like stopping at red lights and carrying on when they are green – which wouldn’t need to be learned.

In driverless cars, the symbol-based transparency of such a hybrid is also crucial. “Symbols are a really important aspect of how we explain ourselves and communicate with other people,” says Bryson. Coming legislation in Germany will require algorithms to explain decisions they take in driverless cars. By 2018, European Union citizens may have the right to ask any automated system to account for its decisions.

However, the most startling consequence of a workable hybrid architecture, Bryson points out, is that it could enable machines to convert their representations into reusable symbols – analogous to language or words (see “Conversational skills“).

“This experiment barely scratches the surface of what we believe is possible with this architecture,” Shanahan says.

Conversational skills

You’d be forgiven for thinking computers have language all figured out. Google can translate between tens of tongues, and natural language processing lets us speak to software agents like Siri and Amazon’s Alexa.

But as Siri’s many noted missteps attest, a computer really has no idea what you’re talking about. It breaks your speech down, gloms on to keywords and makes a good guess at what you’re asking.

For a machine to carry on a real conversation, it must understand what you’re telling it. That’s a much higher-order problem, says Joanna Bryson at the University of Bath, UK, requiring an ability to understand symbols and meanings.

The power to fluidly describe, understand and interact with the world would bring us close to artificial general intelligence, something broadly acknowledged to be a distant prospect. Hybrid systems like the one being developed at Imperial College London may point to a way forward (see main story).