Efemarai

Platform for testing and improving ML continuously. It lets ML teams develop robust models by continuously testing for edge cases.

[Efemarai @efemarai]

While there is lots of great research on the topic, we are really excited to enable ML engineers to use it and get results in seconds!

[Efemarai @efemarai]

The larger the model, the better... right? Wrong! Robustness behaves non-linearly as we show in this comparison between YOLOv8_n (3.2M params), YOLOv8_m (25.9M params) and YOLOv8_x (68.2M params). Very large or very small models are more vulnerable compared to medium sized ones.

[Efemarai @efemarai]

Model comparison is a key aspect of the ML development workflow that is dangerously over-simplified in practice. To choose the best model one needs to consider model properties beyond its performance on static test sets.

[Efemarai @efemarai]

Can you break #YOLO8? Tweak the image and see what the model says. Explore the vulnerabilities of this incredible model and share your highest score!

https://t.co/YfeDqVTREI

[Efemarai @efemarai]

Excited to share our extensive overview of the current ML testing landscape! https://t.co/rMcjw2CQUJ

We hope this structured presentation of the available methods and tools will make the topic more accessible and more teams will look into improving their QA processes.

[Efemarai @efemarai]

We had an amazing time at the INSAIT conference!

The talks about state-of-the-art Machine Learning, Trustworthiness, Verification, Security (or how that has failed in the real and cyber space), NLP inspired us about the amazing work done by great universities and companies!

[Efemarai @efemarai]

Check out our updated Industries Tab! We'll highlight project with different users across sectors revolutionized by AI. Today, have a look at how we helped improve state-of-the-art image tagging models on the path to success. 🚀
https://t.co/0YE7zNspXi

[Efemarai @efemarai]

"Meanwhile, AI systems across industries are susceptible to failures, with nearly a quarter of respondents stating that their organization has experienced issues"

Let's collaborate and develop your AI systems in a robust, secure, and outcome-driven way!

https://t.co/hsakJwI3dn

[Efemarai @efemarai]

We were at Webit last week, met some great people and discussed the future of AI safety with @quaesita! Not only do we need to test everything in order to build trustworthy machine learning systems, but we also must not allow for human negligence at scale. 🚀

[Efemarai @efemarai]

Tell us a bit more about your task, what’s your model based on and the type of dataset you work with, and we will generate for your task:
• detailed robustness and domain coverage scorecard
• 1000 new, meaningful and labelled images containing failure edge cases

[Efemarai @efemarai]

Working in Computer Vision? We will help you improve the robustness and generalization capabilities of your models by discovering edge cases before deploying!

It takes 2 minutes to submit a test request!
https://t.co/Bd78ewUeH5

[TheRecursive @TheRecursiveEU]

Machine learning operations startup Efemarai announced closing a pre-seed round, led by Bulgarian BrightCap Ventures. The solution of the team helps researchers discover, visualize, and fix mistakes in machine learning algorithms.
https://t.co/ChL5oLTlGJ

[Efemarai @efemarai]

Extracting the computational graph, loss function and testing for common mistakes does take just two lines of code. Don't be in doubt, scan by default.
#Machine Learning #Debugging #MLtesting

[Efemarai @efemarai]

The composition of kernel operators allows for progressively more complex symbols to be described. It is interesting to note that with natural images, those first layers kernels look suspiciously like Gabor filters at different target frequencies.

[Efemarai @efemarai]

Understanding the neural network computational graph doesn't have to be hard.
Look for inspiration here: https://t.co/DEuSTTRgZC
#ML #debugging

[Efemarai @efemarai]

When asking for help debugging machine learning code, don't give people a random python file with your model. Let them visually navigate what you have defined or simply share a link with them https://t.co/Jar8bjtBvb.

[Efemarai @efemarai]

How does time sequence data look like in Efemarai? You can clearly see the pattern and distribution!
#MachineLearning #TimeSeries

[Efemarai @efemarai]

Save a particular state of the scan with the new interface. You or a friend will be able to load the file at a later point and get back to debugging where the issues occurred!

[Efemarai @efemarai]

💡⚠️💡 Tips and Tricks Alert 💡⚠️💡
In complex models like Visual Transformers, you can extract a particular feature maps - e.g. the created 16x16 embeddings by using Ctrl+Left click and inspect them in isolation.
#MachineLearning #Debugging

[Efemarai @efemarai]

We support your favorite style transfer models! Look at this lovely bear! #MachineLearningArt