Sometime in the summer of 2012 it was official: the internet is full of cats.
That fact didn't come as a shock to many of us - what did was the source of the research. It wasn't a scientist who drew this conclusion. It was a machine.
More specifically, it was a computer network - fueled by 16,000 processors - that powered the Google Brain project.
Founded by researcher Andrew Ng in 2011, the project used deep learning algorithms to create a neural network with more than one billion connections.
This machine did something quite remarkable. After viewing and digesting approximately 10 million images, it taught itself to recognise cats.
This might not sound particularly impressive on the face of it, but it was a seminal moment for the field of deep learning. Deep learning is a segment of artificial intelligence (AI) and a subcategory of machine learning. Today, it's one of the most rapidly advancing, and fascinating, areas of computer science.
Scientists and industry alike are excited about deep learning because it's mostly "unsupervised" work.
That is, it allows computers to perform tasks for which they haven't been specifically programmed. Parallels are often drawn with the workings of the human brain because deep learning uses artificial neural networks that allow computers to learn and, eventually, "think" by themselves.
Decades of sci-fi speculation have encouraged us to imagine intelligent machines running amok in spectacular ways. But the truth is that computers powered by deep learning are already part of our everyday lives. From canny marketing that knows what music we like, to cars that understand the world around them, deep learning is all around us and it's here to stay.
It's what identifies the voice commands you mumble at Siri. It's what recognises faces in the images you post on Facebook or Google+ and it's what underpins Microsoft Research's new Skype tool, which instantly translates from one language to another. Baidu is using deep learning to target ads on its search engine and deep learning is integral to the future of self-driving cars.
And if deep learning can help cars learn, it can help other machines learn as well. We may be many years away from Ultron in Marvel's Avengers, but similarly intelligent -- though far more benign -- robots are on their way.
One entertaining example is the Robobarista in Cornell University's Robot Learning Lab. Combining deep learning algorithms with a crowd-sourced database of everyday skills, the robot is using knowledge from completed tasks to tackle new, similar ones. Such as making the perfect latte. I can hardly wait for its arrival in my local café.
It also turns out that deep learning models are rather good at identifying the complex patterns and characteristics of cybercrime and online fraud. PayPal has been working with deep learning for a few years. It spoke recently about the potential for one day deploying models that can take live data from its system, learn and retrain themselves, in real time.
Beyond coffee and computer fraud, AI start-up Enlitic hopes to use deep learning for something more immediately humanitarian.
Their idea is to create a system along the lines of the tricorder, the multi-purpose sensing, computing and recording device from Star Trek.
Enlitic's device will collate patient data - from test results to images to medical history - and analyse it using deep learning algorithms to reach a diagnosis and suggest treatments.
According to the founder's recent TEDx talk, the majority of the world's population currently has less than one-tenth the number of trained doctors required to deliver adequate healthcare. Enlitic's device and others like it may help solve the monumental medical crisis that so much of the developing world faces.
These impressive examples of deep learning are well worth watchingbut there's still much work to be done in the field. Today's most sophisticated deep learning models are comparable only with the brain of an insect of terms of overall number of neurons.
While deep learning has recently become more accurate than humans at image classification, other tasks are still beyond it. For example, teaching computers to recognise irony is a task that we have yet to tackle, but if machine learning can already teach them how to make bad jokes (What do you get when you cross a fragrance with an actor? Answer: a smell Gibson), maybe we're not as far away as you might think.Suggest a correction