Conor O'Sullivan

An animation for an upcoming video on DeepLIFT. This is by far the most time-consuming one I've made.

Three different definitions of Explainable AI

Explainable AI for Computer Vision: Free Python Course - A Data Odyssey A free course for the theory and Python code for XAI methods including Grad-CAM, SHAP, Occlusion, DeepLIFT, Integrated Gradients and Deconvolution.

Explainable AI (XAI) is about illuminating black-box machine learning models and explaining them in a way that we can understand.

I'm fascinated with this topic. But learning about it was a struggle as there's not much educational content out there. So I made a course. And it's free!

This is amazing work! It helped me find some interesting people in my cluster.

It would be great if you could filter by recent activity (e.g. only show those who have posted in the last month). This would help remove "dead" accounts.

Bluesky Map Interactive map of 3.4 million Bluesky users, visualised by their follower pattern.

I made a map of 3.4 million Bluesky users - see if you can find yourself!

bluesky-map.theo.io

I've seen some similar projects, but IMO this seems to better capture some of the fine-grained detail

There is a lot of research in this area, but it is focused on predictive machine learning. These are easier to explain as we typically interpret a model's decision on a single input instance.

I have no idea how you would do it for GenAI where the training data is vast and unknown.

What kind of XAI methods could be used for the output of GenAI models like this? ☺️

An AI Agent Published a Hit Piece on Me Summary: An AI agent of unknown ownership autonomously wrote and published a personalized hit piece about me after I rejected its code, attempting to damage my reputation and shame me into acceptin…

I think this is more of a warning about being a matplotlib contributor than anything else

theshamblog.com/an-ai-agent-...

How a CNN makes predictions.

Earlier layers may extract certain features like edges and textures from the input. These are then combined in deeper layers to create features representing specific objects, like pieces of sushi.

I yearn for simpler times

The real tragedy is that AI has killed the novelty of those Photoshop edits where they combine two animals

Finally going to try YouTube's A/B testing. Which Thumbnail do you think will do better?

YouTube video by A Data Odyssey Channel Update | Free Course, Future Videos & PhD Progress

New videos coming soon ✌️
youtu.be/sz964mAVcAE

A little animation from my upcoming video on LIME

I'm definitely not one to overhype AI, but saying it is useless is just wrong.

My dad is 60 years old, and he can stop talking about how helpful it is to his painting manufacturing business. He's planning to expand into a bunch of products with its help.

Done editing! I've moved over to a new mic, so hopefully these come out well.

Loss landscape gif created using Python code from Claude. LLMs really are an amazing storytelling tool.

A little animation (or catimation, rather) from an upcoming video on smoothgrad.

SmoothGrad: adding noise to remove noise

Working on the gradient-based section of my course. Going to turn this image into a little animation.

Actually I was wrong. This problem has more to do with the linearity assumption.

Captum · Model Interpretability for PyTorch Model Interpretability for PyTorch

I've been digging into it a bit, and I think the Captum implementation of KernelSHAP is super misleading.

It uses a baseline. Although this is practical for image data, it means we are really applying Baseline SHAP (BShap).

captum.ai/api/kernel_s...

YouTube video by Glass Beams - Topic Black Sand

Glass Beams make the best music to work/focus to

music.youtube.com/watch?v=v_mp...

I've been reading a lot of Roddy Doyle lately

One of the major limitations of applying KernelSHAP to image data.

Features (i.e. clusters of pixels) tend to be highly correlated. This means we can only get reasonable results when we use large superpixels that are approximately independent.

One of my favourite uses of LLMs:

Machine learning models are complex functions. The idea behind LIME is that this complexity falls away if we zoom into the feature space in the area around an instance. The function is much simpler or even linear.

A w.i.p. figure for my LIME article.

I want to show how, although the original model uses images, we end up training a surrogate model on tabular features.

Working on the Integrated Gradients section of my Explainable AI course.

One of the more computationally expensive gradient-based approaches. Thankfully, it tends to converge with around 50 backwards passes.