Artificial Intelligence and Machine Learning are the next ‘big thing’ – some at Google believe it will have an impact as big as, if not bigger than, the internet.
One of the most important tools of machine learning is ‘deep learning’ – where existing data is fed into a ‘neural network’ of computers that act like a brain (or how we think it acts!) to understand the data and learn from it, with the aim of solving a problem or producing new information. This is repeated over and over again so that the network learns what leads to failure and what leads to success, eventually producing solutions more consistently. Whilst the concept (and examples) of Machine Learning have been around for decades it is only now with the processing power available and the improvement in networking abilities that it is starting to show it’s true potential.
Rather than bore on about the technicalities and potential applications for this technology I’m going to point you at number of existing examples of neural networks and how they are being tested and used currently – nearly all of which are frivolous and highlight how this is a system very much in its infancy.
Learning to walk
Google’s DeepMind division are pretty much the industry leaders for exploring the possibilities. One of their most recent outputs has been getting the system to learn to walk. It’s weird to be rooting so much for a computer generated being that wasn’t produced by Pixar but here we are. This video is adorable.
Learning through play
Probably the most famous example of Deep Learning is AlphaGo which is also from Google’s DeepMind division – this network was able to beat an 18 time world title winner at the game Go, a game that is considered significantly more complex then Chess. During the course of the game the system made a number of surprising decisions that helped it win – it went against the perceived wisdom of humans because it had taught itself how to win rather than relying on the data input (from games played by humans!). In May AlphaGo went on to beat the number 1 ranked player in the world comprehensively and even a team of 5 top players. After the event the DeepMind division decided that Alpha Go should retire – it had conquered all comers and had given human players a greater understanding of the possibilities in Go, but also the possibilities of Artificial Intelligence.
One of the most interesting blogs at the moment on this topic is Postcards from the Frontiers of Science – written by Janelle Shane it documents her experiments with machine learning and neural networks. One of her most popular posts was her attempt to get a neural network to produce new colours and names from a paint companies colour chart. There are a series of posts that outline how the machine learned, how it struggled to associate a colour with it’s name (leading to blue colours being named ‘conk green’ etc.) but eventually with some modifications to how it processed the data it was able to understand what green was, what blue was etc. It didn’t, however, come up with colour names you’re likely to see on the next Dulux chart with names like Stanky Bean (a dull pink), Bank Butt (again a dull pink) and Turdly (rather aptly, a brown colour). Given even more data from different sources the network learned even more and was able to produce colours that could feasibly be seen on your next colour chart – Malora Gray (a light gray), Sinderet Green (a chalky green colour).
An expanding vocabulary
Janelle Shane has rapidly become one of my favourite women in STEM, particularly as her work with machine learning is a hobby (she’s actually a research scientist in optics)! Also because she created the network that produced the ‘Fiat Coma’ as a viable car name – whilst it’s unlikely to get the nod from Fiat in the future I can attest it’s a pretty accurate description of my experience driving a Punto. Another of Janelle’s outputs was to get a network to produce new recipes for the world – some of them have been illustrated wonderfully by Jodee Rose – my favourite result was the ‘Completely Meat Circle’. Whilst it’s been illustrated as a cake by Jodee I can’t help feel the name belongs on a Pizza Hut menu.
Making people laugh
The final one of Janelle’s work i’m going to share is her attempts to get a neural network to produce Knock Knock jokes – purely because this is perhaps the only one where the network was creative and actually produced a viable result. This joke wasn’t in the data that was input so the system learned the basic premise of a Knock Knock joke and was able to produce the following:
Alec- Knock Knock jokes.
You only need to look at Bad Kids Jokes to understand that this is pretty much the beginning of the machine uprising!
There are a number of people exploring the use of neural networks, AI and machine learning in the field of music. You’ll have heard talk of the ‘Hit Machine’ in modern music to refer to particularly succesful writers and producers – ones who have hit on a formula that works. The question is can we actually produce a real hit machine – one that can look at all the data of music and produce the perfect song? There’s an informative overview available on The Conversation about the processes being used in this area. The future of music will be produced by machines…maybe.
This is purely here to show off the fine work of the EMC team. This tool matches up line drawings into photo-realistic images of cats. You can see our efforts above! If you’d like to make your own attempt you can do so here – feel free to tweet us with your creations. Whilst it’s just a fun and interesting tool at the moment it’s easy to imagine rapid prototyping could be done with a tool like this. A quick sketch onto the computer and a photo-realistic model could be produced in seconds.
What can we learn from machine learning?
With the examples above it’s difficult to see how these systems can be a threat to anyone given their current usage. Looking at the gallery above it’s obviously going to be a while before a machine can accurately translate obvious drawings of cats into actual images of cats(I may be overstating our team’s ability to draw cats here!)
There are of course plenty of people putting these systems and data to serious use, however the examples above are accessible and relatable whereas someone creating a system to do complex work isn’t easily explainable and probably doesn’t have results that can be shared at this point. To give a basic example, IBM’s Watson has already been used to successfully diagnose a rare type of leukemia in a patient in 10 minutes when human researchers would have taken at least 2 weeks. The same AI has also been responsible for creating new flavour combinations that are certainly a step above the work Janelle Shane’s network produced!
We need to remember this is just the beginning of machine learning. We’re very much in the ‘baby’ years, it’s just starting to babble to itself and understand the world around it. Much like a child the importance of what we feed it, teach it and how we behave around it will shape it’s future. Look at Google Flu Trends to see how easily we can make false assumptions about the data we’re putting in and the ability of an algorithm or machine learning to make sense of it accurately. This project had the potential to kill the technology at Google before it had even begun!
However done properly and using accurate data it won’t be long before data-driven roles become replaced by more accurate machines – machines that remember everything, don’t miss information and all of the other human ‘failings’.
As an example a role like Legal Secretary is under threat as this is largely a data driven role – the key to survival in the legal profession will be stories and emotional connection, lawyers who bank on technicalities to win cases will become redundant whilst lawyers who excel at outlining the emotional journey will be able to produce work that machine learning can’t better. By utilising what the machine excels best at (data) and what human’s excel best at (emotions) then it’s easy to see how AI can advance human purpose rather than replace it.
I believe we will see this process reflected within the creative industries, particularly marketing and advertising. Realistically the ‘bad advertising’ explained in this article could be produced through automation already. Give machine learning a few years to develop and it won’t just be the production that is automated but also the idea generation. There is a situation in a few years time where a machine could produce the current marketing materials for the new Google phone for example (e.g. analyse the data, see what the new features are, highlight them and stick the product name at the top with a big product image) without any input from a human other than the initial data.
What the machine won’t replace is the more emotional aspect of the current marketing materials – machine learning won’t be able to produce the script for Jonny Ive to read over manufacturing shots, a machine can’t (yet) get that emotional context required for it to be authentic (although there is the debate that it feels inauthentic already!) It’s not enough to rely on technical ability or understanding design science – this is all learnable, understandable by a computer and replaceable. What is important is emotional connection, understanding when and how to tell the story, not just what the story should be. Using the machine to take care of the formulaic tasks of production will allow the creative team to produce more emotional, more exciting campaigns and ones that connect better with the audience.
Whilst these tools will see refinement in the coming years I suspect that they will continue to be restricted to a specific area of expertise. The retirement of Alpha Go highlights how the skills the machine has learnt aren’t easily transferable – it is the undoubted grandmaster of Go but ask it to play bridge or do something outside of the game playing world and much of the information the machine has learned becomes useless. What is important to recognise in this scenario is that it’s the people who created Alpha Go who are moving on to other projects, developing new systems and doing innovative work.
Whilst Artificial Intelligence and Machine Learning are going to replace narrow channels of expertise it’s important to focus on what they offer as a whole. The potential to improve our existence is there, the opportunities for us to automate tasks in our workflow are there and the potential to use these systems to expand our own opportunities is phenomenal. It’s important that we use these networks creatively and use them to further human innovation where possible.
It looks like the key to survival in the creative industries is to be emotionally connected and be innovative with technoology – something that can be difficult to achieve for both humans and computers!