You hear a lot of wildly conflicting stuff about Artificial Intelligence (AI) these days.
On the one hand, there’s massive hype about the “AI revolution”. According to this, artificial intelligence will save the planet and solve humanity’s problems by around this time next week.
And then of course there’s the doom-mongering. Robots will take our jobs, enslave humanity….all the dystopian sci-fi stuff that’s been drilled into us by Hollywood. And then enthusiastically taken up by weirdos on the internet.
Media coverage of AI tends to be sensationalist, because hey, it’s the media.
So sometimes it’s a good idea to take a step back and look at the facts (unfashionable though they may be at the moment).
Where is AI at right now?
We currently live in the age of “weak AI” — artificial intelligence focused on one narrow task. Strong AI, also known as “general artificial intelligence” — a machine capable of performing a range of intellectual tasks at least as well as a human — is thought to be achievable within 10 to 20 years.
If you’ve ever used Siri or Alexa, listened to Spotify recommendations or got a fraud notification from your bank, you’re using AI already. For businesses and organisations, there are a dizzying number of AI applications, from customer service chatbots to financial market analysis.
All of the above employ forms of machine learning. Its simplest version, supervised machine learning, is basically an advanced form of statistics in which humans feed a computer algorithm with training data to enable it to infer things about other data. The more you train an algorithm, the better its predictions get.
Natural language processing
One application of machine learning is natural language processing (NLP) — training computers to understand and generate human (“natural”) language. Algorithms can be trained to “read” text, extract its meaning and arrive at useful insights about it, including what sort of sentiments and emotions it contains. NLP is already being used in a number of industries, from market research to programmatic advertising. At Loyal, we’re using it to scan social media for assets that are useful for journalists and content creators.
As well as machine learning, we’re using a more advanced AI technique known as deep learning. This uses “neural networks” which, as their name suggests, mimic the structure of the human brain and offer far more exciting and often unpredictable outcomes than “linear” algorithms trained by humans. Deep learning can also be unsupervised, which essentially means that the machine can learn by itself.
Remember DeepMind, the London-based company that caused a sensation in 2016 when its human-trained computer beat Korean grandmaster Lee Sedol at Go, a fiendishly complicated board game? Well guess what — a year ago the company unveiled a new version which uses unsupervised machine learning, or deep learning. It beat the previous version 100-0.
The reason we’re exploring deep learning techniques is due to the sheer quantity of data we’re dealing with. Traditional NLP machine learning requires a time-consuming process called feature engineering. The more data you have to sift through, the longer it takes. And with social media, that means a loooong time. To get round this we’re experimenting with catchily-named, state-of-the-art deep learning models like Recursive Neural Networks and Generative Adversarial Networks. Perhaps that’s enough on those for now.
So that’s where we’re at: the cutting edge of an exciting period in AI’s adolescence.