Alberto Rizzoli

Errors in Training Data: How to Spot Them and Their Effect on Model Performance

Misclassified objects, loose bounding boxes, overlapping mask classes - How badly do they affect your AI? Lyft’s Level 5 dataset was found to have missing objects in 70% of its data, some ImageNet classes are up to 92% wrong, and 8 out of 10 ML teams change their label schema between their first model and their production release. How do we stay on top of training data errors, and how do they affect AI deployments in enterprise? We’ll explore examples of how bad training data led to incorrect business results, how to spot errors in your datasets, and how to fix them. This talk will cater to both business and technical audiences, showcasing both qualitative and quantitative results of introducing “bad data” into various computer vision domains.

Alberto Rizzoli is co-Founder and CEO of V7, a platform for deep learning teams to manage training data workflows and create image recognition AI. V7 is used by over 300 global AI companies and enterprises including GE, Fujifilm, Merck, and MIT.

Alberto founded his first startup at age 19 becoming MakerFaire’s 20under20. In 2015 founded Aipoly with Simon Edwardsson the first engine capable of running large deep neural networks on smartphones, leading to the creation of an app enabling the blind identify 5,000 objects through their phone camera used over 3 billion times.

Today he leads V7, one of the UK's fastest growing startups powering the computer vision of millions of healthcare devices, robots, and self-driving cars.

Alberto's work on AI granted him an award and personal audience by Italian President Sergio Mattarella, as well as Italy’s Premio Gentile for Science and Innovation. V7's underlying technology won the CES Best of Innovation in 2017 and 2018.

This website uses cookies to ensure you get the best experience. Learn more