The Illustrator-Moment for Machine Learning?

Cleary, there is a lot going on the Machine Learning area: Day in and day out, new research papers are published, new use cases emerge, tools are more and more driven by the increasing capabilities of Machine Learning. But a lot of this happens behind closed doors. Not just inside of companies that value secrecy but also in academic circles whose terminologies are barely (if at all) understood by the rest of the world. This disconnect between the common not-understanding and the sheer importance of Machine Learning is striking. Machine Learning is shaking up the society in every way thinkable, but only a researcher elite understands it. And, to be honest, even these people hardly understand, what “their” machines are doing.

But it seems, as if Machine Learning as a technology framework is settling in. The underlying concepts are very stable now and with that a new trend emerges: Bringing Machine Learning into the hands of at least a regular computer-savvy user. Two examples:

• CognitiveScale offers a drag and drop interface to put Machine Learning skills and training data together. In a graphical interface users can connect the different elements and build a Machine Learning solution. The company also offers an integrated app store to use models or skills from the community. A dashboard is available to monitor the learning progress.
• Google does something similar with Cloud AutoML, addressing developers “with limited Machine Learning expertise”. Google is rolling out this service gradually but the goal is again to provide a GUI to build, train and deploy Machine Learning models.

This is already going on for a while, but now a new player entered the field: Lobe is also a web application to build, train and deploy models. Lobe, which is currently in invitation-only beta, stands out with its gorgeous graphic interface, and its promised ease of use. I strongly recommend watching the video tutorial to get a grasp of the platforms capabilities:

This is very impressive stuff. In Lobe, you can drag in a folder with your training data and a model is automatically populated from it. With such a foundation, you can start tweaking the model by adding additional layers and fine tune the configuration. Once training is completed, you can use the model and export it to CoreML or TensorFlow, Apples and Googles standards for Machine Learning models. But not just that: Lobe offers a REST API than can basically be plugged into every website and application.

This means nothing less than enabling every designer and developer to build a Machine Learning solution. In his blog article about Lobe John Gruber drew an interesting analogy to the history of computer graphics: When PostScript was released, a lot of people learned to write PostScript to get something printed on paper. In hindsight, even writing such a sentence looks ludicrous, let alone really doing it. But it happened. Then, Illustrator had its appearance, offering a nice GUI for drawing graphics. This was great, but Illustrator also wrote the PostScript code for the user—resulting in a simple “Print” function that took care of everything. Suddenly, learning to write PostScript was no longer necessary. Of course, there was a transition period, where PostScript offered more than Illustrator could do in its interface. But this advantage faded away and writing PostScript became something only printer driver engineers did. Is Lobe to Machine Learning what Illustrator was to PostScript? I’m not making any predictions and the quality of the tool isn’t the only factor to succeed but clearly the market has a desire for Machine Learning to become more accessible. What I find particularly interesting about Lobe: It seems to be not just a nice tool for five or six use case cases and nothing more. Lobe also offers access to the low-level functions of the Machine Learning libraries within its graphic interface. So, like with Illustrator, users can go deep in their design process without losing the convenience of the tool. For me, this is what differs a good-looking proof of concept from a well-designed product architecture.

With that Machine Learning becomes fun—amazing news for every designer.

It’s hard to underestimate the impact of this to the average designer: Since a couple of years I’m a huge advocate of designers to learn something like JavaScript. With design becoming more and more driven by algorithms, I think this is the right thing to do, to control the next generations of design systems (but this is something for a different article). But, honestly, this is not enough to get it going with Machine Learning. Knowing JavaScript is by far not enough to do Machine Learning from ground up. Lobe changes this. Every designer can create simple models with it, and everyone who is willing to learn a bit more, can build unique applications. With that Machine Learning becomes fun—amazing news for every designer. Lobe could help bringing Machine Learning into as many hands possible. Something highly needed to ethically safeguard AIs further progress.

The same happens for developers: No longer the need to learn Python or other languages. Everyone, wo can leverage a REST API (and which developer can’t do that?) can now integrate Machine Learning. (I know, there are other REST APIs to use e.g. image recognition, but these are predefined solutions not your own custom model.) We will see a lot more micro services inside of apps or websites that are driven by Machine Learning. A tool like Lobe also means that designers and developers can collaborate on creating and training the model.

Will Lobe be successful? Will it even be a good tool? I don’t know, maybe a bigger market player will move faster or with more power. But the direction is clear: Machine Learning is about to have its Illustrator-moment.