About me
Hello. My name is Nikola Janjušević. I am a Post Doctoral Researcher at the NYU Langone Health Department of Radiology. I earned my Ph.D. in Electrical Engineering from New York University under the advisory of Professor Yao Wang, NYU VideoLab, in 2024. I received my Bachelors in Electrical Engineering from The Cooper Union for Advancement of Science and Art in 2019.
My current interests are in interpretable deep-learning models for solving noisy imaging inverse problems without ground-truth data. My background is in signal-processing, non-smooth convex optimization, and deep-learning. Outside of academia, I go climbing, biking, and skateboarding with my friends.
Email me at nikola at nyu dot edu.
Updates
[October 2026] The preprint version of our Longitudinal MRI Reconstruciton paper is online.
[May 2026] Our ImMAP Paper was accpted to IEEE ICIP 2026! (Did I) see you in Tampere?
Recent blog posts
TLDR: employing the multi-dimensional chain-rule means writing matrix-multiplication. ...
Convolutional Neural Networks' building blocks aren't just performing the convolution you learned in DSP. In my opinion, the best way to think of these layers is as a channel-wise matrix-vector multiplication of convolutions. 
...
So-called interpretably constructed deep neural networks often sell their methods by showing near state-of-the-art performance for only a fraction of the parameter count of black-box networks. However, can we consider these fair comparisons when the number of learned parameter counts are not matched? ...
A walkthrough of implementing Total Variation color image denoising in the Julia programming language, starring Fabio and Masa. 
...
I often find my downloads folder filling up with tons of research papers with nondescript (ID) names, such as "1909.05742.pdf". Keeping these PDFs open allows me to keep track of them, but once I close those windows they seem as good as lost. To remedy this, I've written a short Python script employing a wrapper for the arXiv.org API. ...
The iterative soft thresholding algorithm is one of the simplest algorithms for sparse coding (in this case, solving the basis-pursuit denoising functional). Understanding its derivation as a special case of the Proximal Gradient Method is a great introduction into the world of proximal methods. ...
Zathura + latexmk -> :). Latest update: 17th January 2021. ...