Synthetic data is the solution to one of the greatest bottlenecks in developing AI models today: data. Vetted data is hard to attain at the scale needed to develop effective AI models. For autonomous driving vehicles, training the autonomous driving system requires millions of images and videos with context of a driving scene, likes cars, pedestrians, or road infrastructure, with appropriate labeling of the image contents. What seems like a simple task, at scale, is actually gruesome, often outsourced and rarely done correctly, bleeding AI startups of their time and money during research and development. Synthetic data is in fact data, like images or video, albeit it’s created by algorithms and not by people. And for synthetic data to be successfully used for training AI models, it must be super realistic at mimicking the real-world. Fortunately, that’s now possible. Super realistic video driving games are used to train autonomous vehicle systems, and that seems to be just the beginning.
DataGen, a newly established Israeli AI startup, has tightened their grip on the synthetic data world but unlike the typical players in the already unique arena, DataGen uniquely targets human feature interactions.
Ofir Chakon, co-founder and CTO of DataGen, started the company as a project in early 2017, but zeroed in on human-machine and human-human interactions (hand gestures currently) after watching Mark Zuckerberg reveal the Oculus Rift platform in October 2017. While the rest of the world cheered for the final debut of major advancement in virtual reality, Ofir turned his attention to one small detail: the Oculus was based on using controllers to maneuver. In anticipation of the transition to natural hand-gesture controlled digital interactions, Ofir foresaw the need for high quality hand-based gesture datasets with 3D hand information and focused his efforts on producing photo-realistic data that manifests the intricate nature of human tissue.
With the understanding that the next wave of computational devices will be controlled by natural hand gestures and movements, Chakon went on to build synthetic hands in a minimum viable product, and diligently build the first version of DataGen’s “Data Generation Engine,” which in turn, created their first large-scale dataset. DataGen fabricates synthetic photo-realistic media designed to train neural networks in order to understand the content of images, video and 3D scanning devices. Using a wide range of algorithms including General Adversarial Networks, DataGen can layer enough realism onto their images to make synthetic data work in the real world.
While autonomous vehicles is the most popular sector for synthetic data today, DataGen is tackling the harder feat: human machine interaction. Human data is much harder to synthesize than automotive vehicle data due to the intricate variations between different people, the soft-body deformations, the complex interactions with the environment and complex simulated motions. Whereas cars have a finite number of unique models and are rigid objects, meaning they keep their shape and don’t deform when interacting with the road. Non-rigid objects like hands, composed of human tissue, manifests its deforming nature from the simple motion of touching an object like a table, affecting its form and even color. Both the non-rigid properties and the interaction between bodies are extremely challenging to synthesize. They’re also super specific, a sub-domain where a huge amount of robotic motion and physical simulation algorithms are currently being developed.
The goal is to fabricate data that is so realistic that the algorithms trained to understand it will also seamlessly understand the real-world. In addition to the visual data, the AI algorithm needs Ground Truth labels providing the visual data with its semantic meaning, explaining to the algorithm what it is supposed to learn to understand. There are many types of Ground Truth labels, each one answering a unique question. For example; Ground Truth labels can be a textual classification label, answering; what type of car is this? Or a segmentation image, answering; what type of object is shown in each pixel? These labels, when generated by hand, are inaccurate, costly, slow and tedious to create. But with synthetic data, the computer generates them perfectly, at scale. This new approach even allows for the creation of 3D Ground Truth labels that were once considered impossible.
The business of synthetic data is not a one-size-fits-all model. They customize the data pipeline for each client and sell them data for each of their products, domain specific. As each project is completed, the team has successfully tailored the synthetic data to an additional domain, using that insight to constantly self-improve. This enables DataGen to grow at lightning speed technologically.
DataGen is fully bootstrapped – no investors and financed with their own revenues. They’re looking to hire genius computer vision researchers and algorithms engineers to expand their brain team and enter growth mode.
Human-machine interaction is just the first tier for DataGen’s offering. Their direction to move from hand to full-body photo-realistic synthetic data is where they see an incredible opportunity ahead. Fast-forward into the future, this might just be the first step to training an Artificial General Intelligence to understand the world around it, from within DataGen’s synthetic world.