Profile avatar
conorosullyds.bsky.social
PhD in ML for coastal monitoring 🌊 South African 🇿🇦 living in Dublin 🇮🇪 I post content about XAI & remote sensing
84 posts 406 followers 793 following
Prolific Poster
Conversation Starter

A few examples of saliency maps used to explain the predictions of deep learning models:

A taxonomy of XAI methods for deep learning

My new post on TDS. Very glad to be contibuting to this great publication again :)

I've fallen in love with QGIS

New cover art for my article on occlusion

YouTube is teasing me! I've had a few videos get this little spike in views and then plateau. Probably for the best, as I'm not really trying to make viral content.

Heatmaps are great, but sometimes you need a more detailed explanation of a model. This is where Guided Backpropagation comes in. See how a simple trick, called ReLU masking, can reveal the complex features used by a deep learning model. youtu.be/C2otM1jHN8s

I published the latest section of my course. Check it out if you want to learn to interpret computer vision models by visualising their gradients of the input image and intermediate layers. The video version will be out next week :) adataodyssey.com/guided-backp...

The Grad-CAM section of my course is finally completed. You can find the article version and video lessons for both the theory and Python application. adataodyssey.com/grad-cam/

The second part of my Grad-CAM tutorial. Check it out for a more hands-on approach :) youtu.be/9NtEMwzPDZ4

New vid hit the top spot 😁 people want to learn about XAI for CV!

label studio needs to add edge polylines setup

In computer vision, heatmaps show which pixels in an image are used by a model to make a decision. Due to its speed, flexibility and reliability, Grad-CAM has become the go-to way for creating these... youtu.be/_QiebC9WxOc

The thumbnail for my next YT vid. Out tomorrow :)

A new animation from my video on Grad-CAM. It shows how feature maps are weighted and summed to create a heatmap.

Editing some YouTube videos and you know what that means… A bunch of cheesy headshots for the thumbnails

Tried to do some filming today. Didn’t go so well! Still working out some problems with my new 24mm lens 😅

I ended up printing a few versions of these maps. So happy they came out well 😁

Thinking of abandoning my current trajectory and becoming a carpenter (for the 14th time this month)

What do you think of the cover for my latest course?

New plot from my article on occlusion :)

The north-west coast of Ireland (source: sentinel-2) Proud to say I can name about 10 surf spots in this image 😄

On a bit of an ML Ops tangent today. Learning how to use hugging face to host models.

Things are heating up in the DL labs. We're implementing multilayer feedforward NNs from scratch. So, I created this diagram to help visualise the flow of activations and gradients during backprop.

My life changed forever when I discovered smoked paprika

The latest addition to my XAI for CV course: adataodyssey.com/class-activa...

I've posted a lot of content. The replies that get me the most worked up is not hate. It's the ones that go: "Hi, I loved your article. Based on what I read, you will find mine useful<link to completely unrelated work> "

I’m teaching the labs for deep learning this year. It’s about that time when attendance starts dropping off.

Story of my life 😂

The input into VGG16 and it's gradients visualised with Guided-backprop

So grateful for @abeba.bsky.social on stage at the AI Action Summit, challenging us to be honest about the fact that we’re moving backward, not forward toward a shared social contract as Big Tech have now walked back nearly every voluntary commitment they once made to be responsible social actors.

I think 2025 would be redeemed if they made a second season of Scavengers Reign

🌌🛰️🔭Wanna know which features are universal vs unique in your models and how to find them? Excited to share our preprint: "Universal Sparse Autoencoders: Interpretable Cross-Model Concept Alignment"! arxiv.org/abs/2502.03714 (1/9)

A diagram that helps explain what's going on with PyTorch backwards hooks. Using it in an article about Guided Backpropagation :)

I love these stories. These researchers are a like a modern day Robinhood 😂