abidlabs

Fine-tuning networks can certainly address some of these issues.

This isn't a research paper describing a novel contribution! Just a simple blog post with some observations I made with the library :)

(Original blogger here.) Yes ImageNet-A is nice! I took a look at the dataset and even referenced an example in the blog post. However, I found that many of the examples are hard to discern myself because many of the images are objects on a camouflaged background, and then many of the images are from rare classes (like "marmots").

I really love the beautiful UX experience

I strongly believe that this should become some kind of standard

As one of the creators of the library, I'll add that we hope to solve one of the biggest problems in machine learning right now:

- ML models are often trained on very specific datasets and fail in embarrassing/costly ways when deployed in real-world settings where things looks different than the training data

- The worst part is that these limitations are usually not detected until the model is released, because usually the ML engineers may not realize that their model is: biased, unfair, or fragile

It takes a diverse team of users and testers to realize this. Gradio gives you the ability to instantly create a web interface around your model that you can share with a public link.

Your users or testers or collaborators can (right from their browsers, without having to install any software) try your model, understand its limitations, and point out biases in the model and data.

Good stuff Andrey! Sounds very familiar :)

If anyone wants to play taboo online: https://abidlabs.github.io/infinite-taboo/

Willpower: https://nationaleczema.org/mind-body-techniques-to-stop-scra...

Thanks! Let me know what use case you have in mind. Iterative collaboration was the original case we had in mind -- as it allowed ML devleopers to iterate different versions of medical imaging model with doctors at the Stanford hospital.

As one of the authors of the "Interpretation of Neural Networks is Fragile" paper, I would agree with you.

To a certain extent, saliency maps can be perturbed even with random noise, but the more dramatic attacks (and certainly the targeted attacks, in which we move the saliency map from one region of the image to a specified another region of the image) require carefully-crafted adversarial perturbations.

Deep learning remains somewhat of a mysterious art even for frequent practitioners, because we usually run complex experiments on large datasets, which obscures basic relationships between dataset, hyperparameters, and performance.

The goal of this notebook is to provide some basic intuition of deep neural networks by running very simple experiments on small datasets that help understand trends that occur generally on larger datasets. The experiments are designed to be "atomic" in that they try to test one fundamental aspect of deep learning in a controlled way. Furthermore, the experiments do not require specialized hardware (they all run in a few minutes without GPUs. See the elapsed time, measured on a CPU-only machine).

Hi guys, I wrote this tutorial for the expectation-maximization (EM) algorithm with the aim of making it accessible to someone with only a prereq of basic probability. If you are curious about the EM algorithm (which is a super important algorithm in machine learning), I hope this helps. If you already know the EM algorithm, I would love to get your feedback on how approachable this is or how you'd tweak it to make it better. Thanks!

This is pretty cool, looks like I know some people who didn't negotiate hard enough ;). It would be nice to get a sense of the standard deviations as well for those positions where you have many data points. Is that possible?