Skip to main content
Menu
Visuals
DeepMind Technologies Limited

Checkmate: how we mastered the AlphaZero cover

We often want to highlight important research on the cover that doesn’t have a clear visual solution. The Silver et al. research article on AlphaZero in the 7 December 2018 issue was one of those instances.

N. Cary/Science

In the game between AlphaZero (white) and Stocksh (black), there were several moves that were reasonable for AlphaZero to consider. After 1000 move-sequence simulations, the red moves were rejected, and after 100,000 simulations, AlphaZero chose the blue move over orange.

AlphaZero is a program that went beyond merely beating humans at some of our favorite games—it managed to beat other state-of-the-art programs specializing in three games: Go, shogi, and chess. The opportunity to visualize this breakthrough on the cover left us with a blank slate. How do you visualize something like a computer program that can’t be seen by the human eye? Ones and zeroes? Programming code? Those are hardly visuals that would be engaging for most Science readers. To add to that, how would we incorporate not just one game, but three, in a way that wouldn’t lack visual focus? As Senior Designer at Science, figuring out this problem fell into my hands.

The Visuals team tossed around ideas with the editors, but little did we know, another team was doing the same thing. The authors of the paper are affiliated with DeepMind, a neuroscience-inspired artificial intelligence (AI) company that develops general-purpose learning algorithms and uses them to help benefit the world. When Science editor Jelena Stajic connected me with the DeepMind team, I saw that they were already working toward what we were trying to accomplish. It was also clear that they were familiar with the problems the Science Visuals team tackles on a weekly basis—finding the best way to visualize complicated scientific processes for a general audience.

Here, Adam Cain, the design lead for DeepMind, and I reflect on how we went from stumped to the cover of Science.

Adam: I lead the creative concepting from initial idea all the way through to completion. But ideas and design aren’t the only things we needed: We worked with a really talented 3D artist called Damien Boudot to realize our concepts, and at every step of the way, the actual AlphaZero researchers and engineers were incredibly generous with their time and ideas, helping us go from initial creative sparks to fully-fledged designs. They pushed us, challenged us, and provided amazing support and even better feedback.

We really benefited from their expertise and understanding—one of them even lent us his prized shogi board so we could study how the pieces look in real life. (Thanks Julian!)

For cover artwork, we like to start with the broadest possible range of ideas. We know cover design is unusual—it has to appeal to the general public and be an interesting, eye-catching image even if you don’t know anything about AI or science—but it also has to really speak to the academic community, who are experts in their field. If they look at the cover, will they understand the concept, will it seem true to them and their understanding of the actual paper it’s trying to illustrate?

DeepMind Technologies Limited

Our first thoughts were around trying to find some interesting spaces that were visually rich (drawing on surrealism, optical illusions, and pattern) but also spoke to some key element of the actual research: the fact that AlphaZero needs no prior human knowledge except the rules of each game, the fact that it mastered three different games, or even the idea of self-play, where AlphaZero learned by playing against itself millions of times.

The process starts with the most basic of tools: pen and paper. After sketching a broad range of ideas as thumbnail scamps, we then move into Cinema 4D, our preferred 3D software. At first, we create simple previsualizations to sense-check ideas. The ones that feel most promising are taken through to the final rendered artwork. For the final part of the process, we use Photoshop to color grade the image and add any final polish.

DeepMind Technologies Limited

Chrystal: This was the first set of images that DeepMind presented to us. I was particularly drawn to the cube. The three games were combined into one image in a way that I hadn’t imagined. Initially we focused just on chess, as this was the most recognizable game, and the editors agreed it was the most complex. Using the faces of a cube, the three games have equal play in the image. By featuring chess on the top of the cube, it becomes the center of attention.

DeepMind Technologies Limited

At this point, I could start tackling the layout. We drastically expanded the size of the cube in the frame. This allowed for all three games to be present, but transformed a very quiet cube into a more dynamic presentation that used the full space available for the cover. This included a pop over the Science logo to add some depth, and clear space within the Go board to allow for room to place the accompanying text.

While the concept was strong, something was still missing. The image felt very static and there was no suggestion of AI’s role in this. At first glance, a reader may guess that the research was simply about board games, rather than AI mastering the games. We discussed this with DeepMind and suggested trying to add motion and a change in color to indicate the technology aspect. They came up with the great solution to show AlphaZero considering possible moves.

DeepMind Technologies Limited

Adam: Overlaying the illustration with real data was a really fun bit! Once we had the concept, we realized to bring it to life we needed real AlphaZero game data. To do this, we drew on existing games from the history of AlphaZero and its development (beginning with AlphaGo), including some games from the latest paper. We chose some interesting moments in each game, and then worked with the research team to understand both the next moves, and, in the case of chess, the assessment AlphaZero was making when it looked at that board position. The moves you see on the chessboard are real moves that AlphaZero would consider.

Science

Chrystal: Now we had a beautiful image that was full of data-laced Easter eggs that I knew our readers would love. However, I couldn’t help but feel that something wasn’t quite right. My eye was repeatedly drawn to the cube itself. The wooden texture conveyed a traditional and vintage look to the game boards, while the warm color distracted from the glowing chess pieces that should have been the focus of the image. DeepMind felt that the wooden cube was reminiscent of a classic chess board and would be nostalgic for chess players. But we agreed to disagree. As Adam said, I “understood [Science’s] audience and what the cover really ‘needed,’ and [was able to] provide that final creative push and insight to help get the cover where it finally ended up.” I suggested that a more metallic texture be used for the cube, which gave the image more of a techy feel.

The final cover is a great collaboration between DeepMind and Science. The cover visual serves to draw readers to an important scientific discovery, and as Adam put it, the cover allows for an “amazing opportunity to create work that will help open up the AI conversation to more people.”

Chrystal Smith is the Senior Designer at Science.