This article was originally posted on our company website. Lakera’s developer platform enables ML teams to ship fail-safe computer vision models.
Building computer vision (CV) products is fun and exciting, it’s magical when you get the first demos working and you can see the results with your own eyes. However, it’s also tedious and notoriously difficult to bring computer vision models to production. The exciting phase happens at the beginning of any new project. More often than not, you have little data to train and test models with and rely on pre-trained open-source models to make your first steps.
At that point, your primary focus is probably to get a proof-of-concept (POC) going that demonstrates the performance of your CV model on a small test data set through your favorite metric (e.g. Precision, Recall, mAP, ROC curve, etc). That’s totally understandable and the right thing to do. However, it’s also where it gets dangerous quickly.
The real work begins after the POC; the POC’s primary purpose is to assess the amount of work involved to build a production model, guide what is needed for that and estimate the chance of success. I speak from experience when I say hitting the target metric on a test set is the easy bit. We all know that. Yet, too often an over-reliance on this one, two, or three aggregate metrics has led me to:
- Be overly optimistic about the production performance of my CV model.
- Overpromise on what can be delivered in a certain time frame and waste resources.
- Collect the wrong type of data.
In short, even at the POC stage, we have to do a rigorous assessment of the status quo of our machine learning models and test them properly. What should that involve? It depends on the product, but at the very least should include (in addition to your target evaluation metrics):
- Testing ML models for robustness through stress and unit tests.
- Performing regression testing to ensure we don’t re-introduce bugs.
- Testing the input data for representativity, bias, and label correctness.
- A model evaluation based on metadata to learn precisely where the model fails and where it works.
These processes used to be a lot of work, but with the advent of dedicated software for machine learning testing, like Lakera’s MLTest, these assessments become as easy as training a new model.
The rewards are immediate, and with insights to accelerate development and tedious processes automated, there’s a much higher chance of success in production.
Most AI software projects are stuck in the prototyping trap or underperform in operation.
An analysis by Gartner in 2022 found that “only 53% of projects make it from artificial intelligence (AI) prototypes to production”.
Why is this the case?
Developers don’t pay attention to model robustness.
All too often developers don’t systematically test the robustness of a model or only do it at the end of development. It’s a widespread (mis-)belief that adding a few data augmentations (e.g. Gaussian noise, horizontal flips, etc) will fix all robustness issues.
Data augmentations can fix some issues, but not all. To know which augmentations to add and what data to collect, it’s key to first understand the robustness of your models and test for that as part of your development as early as possible.
The quality of the data is insufficient.
There is never enough of the right data. Never. At the beginning of a new project, this is particularly true, so you may resort to using any open-source datasets you can get your hands on and become creative in other ways to make more training and test data. Unfortunately, the data you end up with will not be exactly representative of your use case, it will likely contain unwanted biases and possibly dangerous correlations that will lead to your model taking shortcuts.
As with robustness, it’s key to test your models and data against these issues as early as possible; it will help you collect the right data going forward and alert you of potential issues in the final model. Certain applications, like MLTest, help to test for data robustness and where you can proactively procure data that is right for your particular application.
Business stakeholders find it difficult to understand the limitations of the product.
We’ve all been there: a product manager, CEO, CTO or similar has asked if your computer vision model will work under certain conditions. What do you say? “Ehm, probably?”. If you are unsure about the performance of your model, stakeholders one level removed from the technology will be lost, and so will be the customers they sell to.
This can only lead to misunderstandings and frustration for all involved. Canceled projects and canceled contracts. Once you continuously evaluate your machine learning models and data across a wide range of metrics and scenarios, you will be in a much better position to communicate the findings too.
Test early and reap the rewards.
Avoid many of the mistakes we made and start testing early. If you want to learn more about how to test machine learning models take a look at this fantastic overview paper. Or read our more computer vision focused guide on ML testing.
Lakera’s MLTest equips every computer vision development team with a world-class testing infrastructure. Our product gives full visibility into your model’s performance and the quality of your data — automatically as part of existing development processes. Sign up to test your model today!
