May 26, 2017 | Written by: Trevor Davis
Categorized: Industry Insights
In my last post (“A time of wonders“) I finished on the idea that “machines like HAL or Her are referred to as super-intelligent or as strong artificial intelligence…and that most experts agree that we are probably 60 years away from being able to create systems like that, but there are many wonders available to us in the meantime.”
On this image you can see how I think artificial intelligence is going to evolve over the next few years. In my view we are entering a rich period where deep learning technologies will impact almost half of all jobs – impact, not eliminate. There will be new and better intelligent agents that will augment human specialists and in some cases replace them. I’ll show an example later of how even creative professions such as fashion design will be affected. To quote from a recent UBS Q-Series® paper ‘How disruptive will the new dawn of artificial intelligence be?’
“don’t fear the robots, but be prepared to change jobs.”
Unlike previous periods of investment in artificial intelligence, much of the forward motion is now driven by large commercial entities that also have access to the data – IBM, Google and Facebook to highlight three. On this chart from CBInsights, for instance, you can see some of the key players and how they are also consolidating the artificial intelligence market through their investments:
A glossary of AI for 2017
But perhaps I am running ahead. There is a lot of jargon in the artificial intelligence world so perhaps it would be helpful if I clarified some of the key terms.
Today artificial intelligence is broken into a number of domains each of which is making machines smarter and more useful. A common term the people use is machine learning: so what is that? Machine learning uses a wide range of statistical and deep learning techniques to process large amounts of data to identify patterns or solve specific problems such as labelling images or making predictions about the weather. What makes machine learning special is that the machine is not specifically programmed or and there are usually no predetermined rules.
At this moment in time there are two broad families of machine learning:
- statistical techniques which you can see on the left-hand side
- those based on neural networks
When people speak about deep learning there really talking about a type of machine learning that uses layers of neural networks designed to deal with increasing levels of abstraction and break problems down. For example, in image recognition a deep learning system might start by identifying a body, then a woman, then a specific woman.
Or to use another example, in a collaboration with Jason Grech, the Australian couture designer, IBM leveraged Watson to understand, reason and learn from hundreds of thousands of images in fashion archives. These them became stimulus material for Jason to create 12 dresses that were seen on the Opening Gala Runway at Melbourne Fashion Week in 2016.
And deep learning is already delivering results in pattern recognition that are better than human beings.
Once upon a time, computer scientists tried to work directly with abstract concepts or symbols directly, much as people do, but this led to a dead end in AI. This is why computer scientists like to refer to these neural network based techniques as sub-symbolic to highlight the fact that there not really handling abstract concepts in the way the human beings do.
Of course artificial intelligence is always on the move and there are significant research programs looking for new techniques based on cybernetics and simulating how a real human brain works. I will talk a little bit more about these later.
Who watches the watchers?
Most forms of machine learning encountered in business are ‘supervised’. This means that you have to show them an example of what good or bad looks like, and they are then able to reapply what they have learnt to new datasets. An example might be identifying consumer segments based on sample of their social media and purchase data.
If there is a lot of data available, then there are algorithms that can learn without supervision or examples. These are typically used to find structures or features in datasets that no human being could ever work their way through. For example, image search and analysing sensor data from a self-driving car. University of Montreal Professor Yoshua Bengio, one of the most eminent deep learning scientists, sees unsupervised learning as the next area for a big breakthrough.
Google’s AlphaGo AI defeats team of five leading Go players
And with the arrival of deep learning techniques we now have what is referred to as reinforcement learning where the algorithms “decide” for themselves how to improve their performance in pursuit of some goal. This is the approach used by AlphaGo to beat the best human go players: the system plays itself and improves its own performance over time with the aim of winning.
Wonders next time
In the next installment I’ll take a look at what constitutes an AI application. I may even include a few demos!
Playlist for this blog
Jim James, In the Moment, from the album ‘Eternally Yours’
Alice Offley at the Victoria
The starlings in the water-bath in my garden