One of machine learning’s promises is to help humans do things faster and more efficiently. Ironically, one of the roadblocks that keeps businesses and independent developers from capitalizing on ML’s capabilities is that it can be time-consuming and difficult to build, train, and deploy models. PerceptiLabs, a two-person Swedish startup, developed a visual drag-and-drop interface to streamline and simplify the entire process.
It’s designed specifically to offload some of the labor a data scientist or developer would usually have to perform, thereby accelerating the process of development. But it also has pragmatic implications for any business or organization struggling with developing ML tools, because in addition to giving a dev team a speed boost, it allows non-technical people to better understand the process and collaborate.
PerceptiLabs CEO and cofounder Martin Isaksson told VentureBeat that the tool “…Makes it easier to debug [ML models] to understand what’s happening, since we have visualizations for basically everything.”
Getting from there to here
PerceptiLabs started out as essentially a one-off job that Spotify asked Isaksson and his partner, Robert Lundberg, to perform. They taught the company’s 25 in-house data scientists how to apply ML to all the data they were sitting on, and they helped the team create two models and push them into production. After that, they went to work on building their platform full time.
They created what was essentially a beta version, which elicited initial interest from major companies like Amazon, Microsoft, Nvidia, and Google, but it lacked a strong user interface. So Isaksson and Lundberg collected some pre-seed money in part to create the visual user interface and improve the user experience. Over the course of a year, alpha and beta testing has helped to refine both.
In addition to providing developers a speedier means of ML modeling, Isaksson spelled out how PerceptiLabs’ platform solves other problems for businesses that are trying to use the technology.
One of them is the age-old compatibility problem. “There’s so many different [machine learning] frameworks today; even if you use TensorFlow, there’s different versions,” he said. When all of a company’s engineers build their models and push them into production, they need to all be on the same version or there will be issues. If everyone in a company uses PerceptiLabs’ platform, Isaksson said, they would avoid that problem.
Isaksson gave VentureBeat a demonstration of how the platform works, showing that it’s clean and simple enough for a non-developer to use. But he was quick to emphasize that this is a development tool for professionals, not just something for novices to tinker with. “Anyone can build the model, but to really understand what has been done, [and] utilize all the features that make PerceptiLabs unique — for that you need to be a data scientist and data engineer, a dev who already knows how to program a model,” he wrote in an earlier email conversation.
But, he said, some of the direct feedback that PerceptiLabs has received from beta testers is that it’s simple enough visually that product managers feel like they can grasp the complex models that the data scientists are creating. That’s a form of transparency, and it can engender a certain peace of mind for managers who otherwise may feel like the development work is a black box.
Overall, Isaksson, noted, PerceptiLabs hopes that this platform lowers the barrier of entry for more people and companies to get involved in machine learning, which would have the dual effect of potentially pulling more people into the field to fill the currently empty empty data science jobs and also giving those people a means of doing their work more efficiently.
There are two versions of PerceptiLabs. The free version is available for anyone to download, with a getting started guide. It has all the visual elements and can run models locally — Isaksson performed the live PerceptiLabs demo for VentureBeat on a laptop while running a simultaneous video chat — but the enterprise version can run on a company’s own hardware or in the cloud.