The creation of AI applications is accelerated by Encord Active, a fully open source active learning toolbox for computer vision.
Encord Active is a free open source industry-neutral toolkit that enables machine learning (ML) engineers and data scientists to assess and enhance the quality of their training data and contribute to improving model performance. Encord is a platform for data-centric computer vision.
AI experiences a “production gap” between effective proof-of-concept models and models that can operate “in the wild” for many use cases, including self-driving cars and diagnostic medical models. Proof-of-concept models work well in research settings but have trouble making reliable forecasts in practical situations. This gap results from problems with model reliability and robustness that have hampered the broad use of AI.
ML engineers can close this gap by utilising a fresh method for evaluating the accuracy of their data, labels, and model performance with Encord’s open source toolbox. For high-quality predictions, it is essential to regularly assess and enhance training datasets because data and label mistakes can significantly affect a model’s performance. The new solution from Encord gives machine learning teams the ability to identify model failure modes, give high-value data top priority for labelling, and promote intelligent data curation to enhance model performance.
Active learning has gained popularity as a theory among researchers, start-ups, and businesses. Active learning is a process for training models in which the model requests data that can help it improve its performance. However, smaller AI firms have not yet been able to put workable active learning approaches into practise. Encord Active offers a new methodology built on “quality measurements,” enabling businesses of all sizes to get from theory to practise. Your data, labels, and models will have computed indexes added on top based on ideas that are easily understood by humans.
Active learning techniques now in use need ML developers to build their own tools and versions of quality measurements, which adds time and cost to the process. By automatically calculating a variety of pre-built quality indicators across the data, labels, and model predictions, Encord Active eliminates that work.
The automatic calculation of picture properties, labels, model predictions, and metadata is the main goal of the quality metrics approach. The data, label distribution, and model performance for each metric are then broken down for ML teams. These understandings enable them to:
- Explore their databases for unidentified failure modes.
- Before labelling or training a model, check if the dataset is balanced across the various metrics and balance it according to the quality metrics.
- Find probable outliers in their dataset that can be eliminated if they are not needed for the use case.
Encord Active is the first platform to offer practical end-to-end active learning workflows, enabling models to continuously learn and advance in a manner akin to that of people. Users can complete the workflow cycle to resolve these issues as well as locate important data to label and label errors to re-label within the Encord ecosystem.
Encord is supported by CRV, Y Combinator, WndrCo, and Crane Venture Partners. It is trusted by world-class healthcare organisations like King’s College London, where it assisted in annotating pre-cancerous polyp videos, increasing efficiency by an average of 6.4x, and automated 97% of labels, making even the most expensive clinician 16x more productive at labelling medical images. Additionally, it has collaborated with Stanford Medical Center and Memorial Sloan Kettering Cancer Center, where it has decreased experiment duration by 80% and processed three times as many photos.