Aquarium helps teams build and improve their ML models.
ML models are defined by a combination of code and the data. While there's a lot of great tools for debugging and understanding code, there's not a lot of tooling for debugging and understanding data. Yet the majority of model improvements come from improvements to the data!
Aquarium makes it easy to find + fix problems in your model performance. Users can easily find bad data and patterns in model failures, and then fix these issues by smartly editing / adding data to their datasets.
Our goal is to make it easy for ML teams to analyze and improve their models, then share their results with other team members.
See our user guide to get started!
You should use Aquarium when you're trying to build or improve an ML model. It's useful for common ML tasks like building your first datasets, finding quality issues with your labels, or finding what data to label next to best improve your model.
Here's some common pain points that Aquarium tries to address:
Your modelβs accuracy measured on the test set is at 80%. You abstractly understand that the model is failing on the remaining 20% and you have no idea why.
Your model does great on your test set but performs disastrously when you deploy it to production and you have no idea why.
You retrain your model on some new data that came in, itβs worse, and you have no idea why.
Hypothetically, Aquarium's technology supports any task where one can utilize a deep learning model.
However, we have first class support for the following:
Modality | Classification | Detection | Semantic Segmentation | Instance Segmentation |
Image | Yes | Yes | Yes | Yes |
3D Pointcloud | Yes | Yes | No | Yes |
Audio | Yes | N/A | Yes | Yes |
NLP | Yes | N/A | N/A | N/A |
Check out our full manifesto from our Hacker News launch post here.
Our website: http://aquariumlearning.com/β
Our Hacker News launch post: https://news.ycombinator.com/item?id=23821502β
Software 2.0: https://medium.com/@karpathy/software-2-0-a64152b37c35β