Synthetic AI startup raises $17M to train facial recognition with digitally-rendered people
Startup Synthesis AI has raised $17 million in Series A funding from a number of funders to back its synthetic AI platform that generates a wide range of people to teach visual AI models like facial recognition.
Synthesis AI, founded and led by CEO Yashar Behzadi, received funding from a round led by 468 Capital with participation from Sorenson Ventures and Strawberry Creek Ventures, Bee Partners, PJC, iRobot Ventures, Boom Capital and Kubera Venture Capital, according to TechCrunch. The proceeds will support its product R&D, team growth, and further research, according to Behzadi. There are plans to develop products in verticals including photo enhancement, teleconferencing, smart homes, and smart assistants.
The company specializes in the generation of synthetic data on a cloud-based platform that creates synthetic image data with labels through a combination of AI, procedural generation, and video effects rendering technologies. It allows for modifications of bodily features like gender, age, weight, skin tone, ethnicities, and more specific appearance details like pose, hair, facial hair, and apparel. There is also adjustment for environmental aspects like the lighting and even the lens type of the virtual camera.
Synthetic data is growing in popularity within the biometrics sphere to train algorithms for its relative ease of use, customization for humans and environment, and potentially avoiding the ethical and legal pitfalls that accompany the collection of data from people.
Behzadi told TechCrunch that AR, VR, and metaverse companies are already using Synthesis AI’s platform to build “more realistic and emotive avatars,” as well as smartphone and consumer device customers adjusting their camera modules, and developers building a driver and occupant sensing system that use the synthetic data to figure out optimal camera placement in a vehicle.
The chief executive believes Synthesis AI can solve problems of gender and racial bias that affect the collection of face biometrics. He holds that the platform can generate a more diverse assortment of faces that would otherwise be uncollected in the real world. It is also said to not carry personally identifiable information and isn’t copyrighted, which adds privacy and fair use.
A research paper from Microsoft demonstrated the possibility of using synthetic data to train facial recognition algorithms, but a 2021 article in the MIT Technology Review suggests mixed results from attempts to bypass privacy and bias allegations with synthetic data. Matthew Guzdial, an assistant computer science professor at the University of Alberta, is quoted as saying in the TechCrunch article that Synthesis AI’s own whitepaper admits that a model trained on solely synthetic data will result in worse outcomes.
The funding of Synthesis AI joins the likes of Datagen which has raised millions for synthetic AI to train AI algorithms in human motion biometrics.