The AI mastermind Andrew Ng announced a data-centric AI competition on Friday. The competition will include improving the performance of machine learning models by optimizing the data rather than the model/algorithm.Â
The data-centric approach is not always an ideal do-to method when it comes to improving any model’s output, but according to a few recent pieces of research that were performed by Andrew’s team, it has shown that there is a significant improvement in the results using data-centric methods rather than model-centric ones. This led to the invention of the competition to know more data-centric approaches.
Users can register for the data-centric competition, and then download the datasets from the website that consists of handwritten roman numbers from 1 to 10. The task is to optimize model performance by improving the datasets via training and validation sets. The various data-centric techniques such as fixing incorrect labels, adding own data or moving trained and validation splits can be implied.
Read more: CSEM Develops Artificial Intelligence Powered Chips That Runs On Solar Energy
The submission has to be done on the Codalab website by uploading a zip file consisting of less than 10,000 png images which work as the datasets. There can be only five submissions per day. The uploaded folder and the label book will be passed on to a predicting script that will train a fixed model on the submitted data. The script will generate a set of predictions on the label book. The competition organizers will mimic the contender’s run by replacing the deb set (label book) with a test set to obtain accurate results. The accuracy will be uploaded on the leaderboard. The results will be based on the best overall performance and the most innovative approach.
The data-centric competition will be open till September 4th. There will be two winners from each category who will get an opportunity for a private discussion with Andrew Ng regarding data-centric optimization. Their work will also be published on the deeplearning.AI channel.
This will be the first-ever data-centric competition, but there will be many more in the coming years Ng mentioned.
