$3 million to clean up data for machine learning

Matthias Heller (L) and Igor Susmelj (R), former ETH Zurich and Harvard students, who worked in autonomous driving and research on computer vision and deep learning, witnessed first-hand how many engineers' efforts are excessively directed into manually cleaning datasets or how data quality is disregarded and the model becomes the focus. Learning that machine learning (ML) is like software that is updated with data instead of code, they developed Lightly in 2020 to drive quality data into ML models to generate better outcomes.

“Most companies that use artificial intelligence need to label data. However, the amount of data produced in a single day cannot be labelled in a million years. That is why it is vital to choose wisely what data to label and put into an ML model,” highlights Matthias Heller, co-founder of Lightly.

The Lightly data curation tool deploys self-learning algorithms that tell data teams which 1% of their data is the most valuable. This approach allows companies to reduce data selection and labelling costs by up to 90% and improve the performance of their AI model by 20%.

Additionally, the tool tackles the challenge of biased data, attempting to make AI more ethical, it powers ML models with recent data ensuring they remain relevant and manages such data changes by intelligently adding novel data to datasets. Pioneering in the field by automating data pipelines, Lightly makes data management easy by integrating other data preparation tools and offering a centralized platform. This supports the trend toward data-centric AI, a concept elaborated by Andrew Ng. In his words: “Data is food for AI”.

Today Lightly is used by Fortune 500 companies and startups in autonomous driving, robotics, and video analytics. To accelerate its growth, the company plans to expand its team and enter the US market this year with a new office in the Bay Area. This ambition will be facilitated by the capital obtained as part of its seed round. Wingman Ventures led the round.

(Press release/RAN)