Creativity: how is AI impacting this human skill?
AI is the product of processing a mixture of data, often aided by other information (labels or domain models, such as Wolfram Language). The big emphasis today is on the former, but there is enormous merit in having a balance of the two.
Creativity is one of the most human qualities. So how is the rapid evolution of AI and machine learning affecting our creativity? Here’s a simple template:
- Clarify: what do we mean by “creative”?
- Illustrate: show examples of both products and project into the future based on research today around AI creativity.
- Act: Where does this leave this human skill? What should we do concretely?
Clarifying: what is creativity?
It’s a tough word to define. Let’s see some definitions:
“Creativity is reorganisation of existing knowledge” — Pierluigi Serraino
“Creativity is the phenomenon of something new and somehow valuable is formed” — Wikipedia
“Creativity is… making a connection between item 1 and item 2 which might not ordinarily seem connected.” — Isaac Asimov
As I’m specifically looking at this human skill from the point of view of automation, I’ve landed with this definition:
Creativity is an assessment of the diversity and complexity of information used to make new things.
Let’s test it out:
- There is no value statement in creativity: I think this is fine. You can be very creative and produce nothing of value.
- It’s a reflection of a process rather than definition: yes. You can have a creative process, one which produces things that are creative, which is said to have creativity in it.
- The assessment criteria is missing: creativity does mean different things in different situations; for the workplace it is usually associated in a positive way with “intuitive leaps” and “good application of imagination”. In creative industries, creativity is more to do with aesthetics.
Illustrate: creativity with AI
I’ve attempted to group (probably not very successfully) two ways in which technology is becoming more creative: prediction, and imitation.
Prediction is can actually appear to be very creative: explicitly it takes past data and uses it to create a prediction for the next piece of data. You can add that prediction to the set and run a new prediction. This is illustrated with astonishing effect in Googler Martin Gorner’s demo of how to create plausible (but nonsensical) Shakespeare using a thing called a “recurrent neural network” or RNN:
At this point I draw out my favourite tech quote: “Any technology sufficiently advanced is indistinguishable from magic” — Arthur C. Clarke
Are there others? Yes indeed. More coherent writing is possible with Narrative Science, Logojoy makes logo creation point-and-click, AlphaGo in its specialised zone was awesomely creative, the list goes on. With sufficient interest I will drill into this area further.
We are at the dawn of a new world of creativity aids that would be seen as mind-blowingly creative compared to even 5 years ago.
Deep learning and GANs
A quick aside: machine learning for images (deep learning) works with a series of layers that build increasingly sophisticated understanding of an image. Andrew Ng’s image is a great explainer:
“Way back” in 2015, some clever fellow asked the question “what happens if I reversed this? What would it produce?” On a personal note I was still at Google at the time and I remember an internal post something to the effect of “this is what happens when you code until 2am in the morning”. This kind of thing:
As the code was open sourced, innovation was abound. Quickly, Prisma came along:
Prisma is an app that takes the style information from one photo and applies it to another. I know I’m on to something when my wife, a trained artist, looks at pictures above and says “I love how it’s done the transparency of the hair”. There are a stack of apps in this space; Oilist is another fun one of note.
Google Autodraw is an experiment that attempts to guess what you’ve drawn and give you a better version. I’d like to say it failed to recognise my miserable attempt at a face due to a limited test data set but I in reality there is a limit to how much you can do with such material!
Create new faces with a single brushstroke
Moving closer to the twilight zone see how this experimental tool uses basic colour choices and pointer gestures to create completely new pictures.
Text to image
“Ok Alexa, draw me a picture of a red and brown bird with a stubby beak”.
Science fiction right? Decide after learning about StackGAN. StackGAN is a research project that takes a phrase and creates a completely new picture that represents it. You’ll notice that a lot of those images look pretty bizarre, but in terms of progress, it doesn’t take much imagination to consider where this is headed:
Act: developing resilient creativity skills
The two key strengths of humans over computers is relatively straightforward: direct access to the real world and adaptability.
Strive to gain a wide range of experiences and expertise
We have great access to the real world. If you think about the difference between “Job title” and “Senior Job title”, it’s usually, at its core, evidence of having creativity in an increasingly complex and broad domain. Computers can only see patterns in data they have, and real world experience over years creates an enormous pool of to draw creatively from. As such, successful careers, more and more, are going to be ones that expose different situations, different problems and and different outcomes.
On the AI horizon, devices are spreading throughout the world to collect data (in the home and in the workplace). Equally, designing better ways to learn faster is also a very active area for AI. Virtual simulation looks particularly promising: the ability to simulate a good approximation of the real world environment of the scenario in question entirely virtually, which then enables testing many thousands of times faster than in the real world.
The mattress salesman can the next day be the washing machine salesman with only a small amount of training. A programmer can pick up a new language in a day or two. You can learn to ride a motorcycle with only small amount of practice after learning to drive a car. Humans are intrinsically adaptable but as a hangover from the industrial age a lot of our professional training actually reduces this. This trend is reversing and, in the face of AI, is key to a resilient career. Specific suggestions:
- Think of adaptability as a habit, something to practice. This means you will probably fail a lot as I have. Accept this, brush yourself off, ask “how can I do this better next time?” and try again. Encourage people around you to help pick you up too.
- Invest time finding a team or environment that demonstrates an adaptable culture. It’s really hard to be adaptable in an environment that is indifferent, or even outright reticent to it.
“Immerse yourself in the community of people being like you want to be” — Paramhansa Yogananda, Man’s Eternal Quest
On the AI horizon, adaptability translates to “transfer learning”: each of the examples I provided here are transferring an unarticulated abstraction of the learning concepts to accelerate learning in a new sphere. While this is still largely an aspirational frontier, it is a reality in some specific areas such as Google Translate, that appears to have developed a way of using what it learnt in one language pair (e.g. French < = > English) to other languages.
If you enjoyed this piece and would like to see it fleshed out more, please recommend it and I’ll use this as an indicator.