It's fixed.

Our paper "Visual Graph Arena: Evaluating AI's Visual Conceptualization" has been accepted at #ICML2025! 🎉🥳🎉. We introduce a dataset testing AI systems' ability to conceptualize graph representations. Available at: Camera ready version coming soon!

Make peer review non-anonymous. Just like in other fields of science.

I noticed the image quality is not great! It should be due to Twitters image compression. Here is a link to the PDF file of the poster:.

If you are interested in generalization in deep learning, you can find our poster in the poster session right now at #AAAI2025 : )

Camera-ready version is out! (. TL;DR: Deep CNN filters may be generalized, not specialized as previously believed. Major update(Fig4): We froze layers from end-to-start instead of start-to-end. The result ironically suggests early layers are specialized!

RT @giffmana: I took a brief look at the Harmonic Loss paper. tl;dr: instead of dot-product with softmax, do euclid dist with normalized 1/….

A company created a huge open-source model along with a full paper. They publicly shared everything they had—unlike closed companies. Now, people are debating whether they are a "threat" or not. I wonder if the reaction would be the same if they weren't Chinese.

RT @PaulGavrikov: Our discoveries? VLMs naturally lean more towards shape, mimicking human perception more closely than we thought! And we….

13/13 If you prefer video content, you can check out the video I made for AAAI:. Thanks for reading! I wish a great day for you. #DeepLearning #ComputerVision #AAAI2025.

12/13 Thank you for reading this far! Curious to hear your thoughts on this. You can find our complete paper "The Master Key Filters Hypothesis" here:

11/13 We went even further, transferring FROZEN❄️ filters between two different architectures (Hornet → ConvNeXt) trained on unrelated datasets (Food → Pets) improved accuracy by 3.1%. We suggest these "master key" filters are truly architecture-independent!

10/13 To test this, we conducted cross-dataset FROZEN❄️ transfer experiments. If true, filters trained on larger datasets should help models perform better on smaller, even unrelated datasets - as they'd better converged to these "master key" filters. Results were confirming.

9/13 We propose that there exist universal "master key" filter sets optimal for visual data processing. The depth-wise filters in DS-CNN naturally converge toward these general-purpose filters, regardless of the specific dataset, task, or architecture!.

8/13 This contradicts the intuitive notion that models should develop more specialized filters as get deeper. Enter our "Master Key Filters" hypothesis. 🔑.

7/13 With top model reaching ~94% retention, the trend suggests near~100% retention may be possible with better models. Conversely, as accuracy drops, we approach Yosinski's findings. But why do deeper, better-performing, models become MORE general, and not more specialized?

6/13 Our investigation says no! When we transfer all of the layers of ResNets and draw a chart of "retained accuracy %" vs "original accuracy", a clear trend emerged: the better a model performs on its original task, the more "general" (not specialized) its deep filters become.

5/13 We had to test with CNNs too. ResNet50 (a classical CNN) maintained 92.5% of its original accuracy when all layers were transferred! This number reaches to 93-94% for ResNet152. But why? is this because of residual connections?

4/13 So we replicated Yosinski's experiment with modern DS-CNNs: we split ImageNet into man-made and natural images, transferred & froze filters between models up to layer 'n'. Surprisingly, there was NO significant accuracy drop!. But why? is this because DS-CNNs architecture?

3/13 Our previous paper, "Unveiling the Unseen", revealed something intriguing: DS-CNN filters showed remarkably similar patterns across ALL layers! This made us question: how can early layers be "general" and deeper ones "specialized"? Time to revisit Yosinski's experiment. 🤔