- Every matrix can be triangularized by choosing an orthonormal basis. - Really, are you "Schur"?

While people have been discussing ethical side of moral decision of driverless car, now doctors have to decide who will have treatment in case of limited resources in reality.

Oversampling should be done after test-train splitting and only for train split. This is also valid for cross-validation. Otherwise that causes 1)data leakage due to duplicated samples, 2)biased performance measure due to test-valid set difference. The same goes for SMOTE.

In Pytorch, recommendation not to use bias=True for linear or conv layers before BatchNorm (BN) is not correct when using ReLU between them. Bias is useless since BN centers the values again but this is not the case when ReLU exists before BN. (Ex: Linear-ReLU-BN)

Self supervised learning is revolutionizing AI using large unlabeled datasets. We show that maximizing mutual information between alternative representations of the same input is a practical method for self supervised learning that is immune to the dreaded collapse problem.

Our paper (arxiv.org/abs/2209.07999) just got accepted to NeurIPS Conference 2022! Congratulations to Alper Erdogan for coming up with the idea and Serdar Ozsoy, Shadi Hamdan, Sercan Arık for a great collaboration.

The existing biologically plausible neural network approaches to solve the blind source separation problem typically assume independence/uncorrelatedness of sources. In our article, we propose an alternative framework that is capable of separating correlated sources.

Our paper (arxiv.org/abs/2209.12894) just got accepted to NeurIPS Conference 2022! Conference. Many thanks to my collaborators Bariscan Bozkurt and Cengiz Pehlevan for this exciting joint work.

We’re excited to announce that TabNet is now available in Vertex AI Tabular Workflows: lnkd.in/g7jB5EMe! Tabular Workflows provides fully managed, optimized, and scalable pipelines, making it easier to use TabNet without worrying about implementation details. (1/3)

📢We are happy to share that KUIS AI members F. Güney and Kaan Akan presented their paper StretchBEV at #ECCV2022. Congrats again👏🏽 check out #ECCV2022 magazine for their interview! rsipvision.com/ECCV2022-Wedne… #computervision

💫We're happy to share that our MS student Serdar Ozsoy and faculty members Alper Erdogan and Deniz Yuret presented posters of the two papers at #NeurIPS2022 in USA, and our MS student Bariscan Bozkurt made a successful virtual presentation in #NeurIPS2022 Oral Session. Congrats!👏🏼

Here are the papers: "Biologically-Plausible Determinant Maximization Neural Networks for Blind Separation of Correlated Sources" #NeurIPS2022 (github.com/bariscanbozkur…)

"Self-Supervised Learning with an Information Maximization Criterion" (github.com/serdarozsoy/co…) #NeurIPS2022

Confused about #NeurIPS2023 rebuttal policy. Can we submit multiple threads for each reviewer? With 6000 characters limitation, it is hard to respond to all reviewer comments/questions. Can we also use "Official Comment" to provide our responses.

🎉Thrilled to announce that our paper for #NeurIPS2023 titled “Correlative Information Maximization: A Biologically Plausible Approach to Supervised Deep Neural Networks without Weight Symmetry” has been accepted! (arxiv.org/abs/2306.04810)

Join us at the poster session for our #Neurips2023 article "Correlative Information Maximization: A Biologically Plausible Approach to Supervised Deep Neural Networks without Weight Symmetry", our joint work with Bariscan Bozkurt & Cengiz Pehlevan. 🗓️ Wed, 5-7 p.m. 📍 Hall B1+B2 #423

Great compilation of his insights on diffussion models. For a deeper dive, head over to his blog. Highly recommended! -> sander.ai

The new content race on social media: DeepSeek.

My new habit, starting with reasoning models: not going to the final answer without reading the chain of thought. Less efficient, more enjoyable.

Have you ever seen a learning curve that looks like a step function? It turns out a few hundred negative examples flips a switch inside an LLM and gives a discrete jump in accuracy. "How much do LLMs learn from negative examples?" (arxiv.org/abs/2503.14391) with Shadi Hamdan.