The Alien Style of Deep Learning Generative Design

Source: http://www.arup.com/news/2015_05_may/11_may_3d_makeover_for_hyper-efficient_metalwork

What happens when you have Deep Learning begin to generate your designs? The commons misconception would be that a machine’s design would look ‘mechanical’ or ‘logical’. However, what we seem to be finding is that they look very organic, in fact they look organic or like an alien biology. Take a look at some of these fascinating designs.

The photo above design is described as follows:

“This is not only an exciting development for the construction sector, but many other industries as well. In the case of this particular piece, the height is approximately half that of one designed for traditional production methods, while the direct weight reduction per node is 75%. On a construction project that means we could be looking at an overall weight reduction of the total structure of more than 40%. But the really exciting part is that this technique can potentially be applied to any industry that uses complex, high quality, metal products.”
— Salomé Galjaard, Team Leader at Arup

This is a car frame that is designed by a generative algorithm.

Source: http://hub-apac.insight.com/h/i/189108757-cad-is-a-lie-generative-design-to-the-rescue/166669

The design on the right of an antenna is twice more effective that that on the left.

Source: http://www.instructables.com/id/How-to-Design-a-Bike-Stem-in-Dreamcatcher/

Above is a lightweight bike stem generated by an algorithm.

Source: http://inthefold.autodesk.com/in_the_fold/2015/07/autodesk-within-generative-design-optimized-for-3d-printing.html

Above is a lightweight engine block.

Many of these designs come from Autodesk’s DreamCatcher research.

The Dreamcatcher system allows designers to input specific design objectives, including functional requirements, material type, manufacturing method, performance criteria, and cost restrictions. Loaded with design requirements, the system then searches a procedurally synthesized design space to evaluate a vast number of generated designs for satisfying the design requirements.

Generative designs also exist not only in the physical world but also in the design of neural networks themselves:

These LSTMs are designed by an algorithm and shown to be more effective that the conventional LSTM. These are generative neural architectures, more like meta meta-models.

What is surprising is that these designs do not exist for the sake of style. Rather, these designs are actually the optimal solutions to multiple competing design requirements. Why do they look organic or biological? Is there some underlying fundamental principle that exists in biological systems that leads to this? Why aren’t the solutions sparse, but rather complex?

Even a more deeper question is, if these were the optimal designs, then why don’t inanimate objects look like this? Inanimate objects that are complex tend to have a fractal style:

Credit: http://www.fhshh.com/the-discovery-of-fractals-extraordinary-examples-of-geometric-complexity-of-mother-nature.html

The self-similar repeating patterns that we see in crystals and coastlines ,despite looking complex, certainly have a style that is different from organic or biological styles. Deep Learning clearly has similar capabilities as biological systems. I suspect that this difference originates from the difference in computational machinery that generates these. It indeed is fascinating that the style of these generated objects are a reflection of the capabilities of its creator.

One final thing though, just because DL exhibits some behavior that appears to be biological, it still is a far cry away from something that is intelligent.

Know more about this at Intuition Machine or join the FaceBook group: https://www.facebook.com/groups/deeplearningpatterns/

Update: Wired writes about a concert hall designed by algorithms: https://www.wired.com/2017/01/happens-algorithms-design-concert-hall-stunning-elbphilharmonie