I’m a Master's student in Mathematics and Computer Science at Saarland University. I'm also a researcher at CISPA working at MLO Lab with Prof. Sebastian Stich. My main focus nowadays is on optimization theory. Also check my Latent CV from here :))
My research interests include but are not limited to the following areas:
So far, my research has primarily focused on optimization theory, particularly in distributed settings and minimax problems. I have worked on understanding the effectiveness of Local SGD from a theoretical perspective, providing a new convergence guarantee for this method. More recently, I have been developing a communication-efficient approach for solving minimax problems in a distributed manner. Lately, my interests have shifted toward deep learning theory, specifically the role of optimizers in both the training and testing phases, with a focus on generalization error. I aim to gain deeper insights into how the implicit bias of modern optimizers affects generalization performance.
I'm always open to meet new people and chatting about new topics in ML. Please let me know if you are interested in working on a project with me. You can simply send me an email.
Currently working on distributed optimization.
Worked on adversarial attacks on self-supervised models.
Worked on the problem of off-road prediction in self-driving cars. I leveraged a contrastive representation learning approach for this problem to improve the latent space and I reduced the off-road prediction rate about 12%.
Developed an algorithm for scanning the cross-section of metallic products and made a 3D visualization of them.
Proposed a bifurcated neural network for segmentation of COVID-19 infected regions in CT images. I also developed a method based on Pix2Pix conditional GAN to generate synthetic data with the goal of data augmentation.
Worked on a cloud computing platform called “Open Stack” to deploy it at the information center of university.
Worked as a software engineer at Sitco company to develop an accounting software.
In this paper, we provide new lower bounds for local SGD under existing first-order data heterogeneity assumptions, showing that these assumptions are insufficient to prove the effectiveness of local update steps.
View PaperHowever, there is a gap between the existing convergence rates for Local SGD and its observed performance on real-world problems. It seems that current rates do not correctly capture the effectiveness Local SGD. We first show that the existing rates for Local SGD in a heterogeneous setting cannot recover the correct rate when the global function is quadratic. Then we first derive a new rate for the case that the global function is a general strongly convex function depending on third-order smoothness and Hessian similarity.
View PaperThis research proposes a method for segmenting infected lung regions in a CT image. For this purpose, a convolutional neural network with an attention mechanism is used to detect infected areas with complex patterns. Attention blocks improve the segmentation accuracy by focusing on informative parts of the image. Furthermore, a generative adversarial network is used to generate synthetic images for data augmentation and expansion of small available datasets. Experimental results show the superiority of the proposed method compared to some existing procedures.
View PaperThis paper proposes an approach to segment lung regions infected by COVID-19 to help cardiologists diagnose the disease more accurately, faster, and more manageable. We propose a bifurcated 2-D model for two types of segmentation. This model uses a shared encoder and a bifurcated connection to two separate decoders. One decoder is for segmenta-tion of the healthy region of the lungs, while the other is for the segmentation of the infected regions. Experiments on publically available images show that the bifurcated structure segments infected regions of the lungs better than state of the art.
View PaperIn this project, we aim to provide convergence rates for different first-order methods used for min-max optimization problems. We considered the methods: Gradient Descent Ascent (GDA), Proximal Point (PP) and Extra Gradient Descent (EG).
View BlogIn this project, we aim to show the effect of vari- ance reduction methods in convex and non-convex optimization problems and we show how they can be better than pure stochas- tic methods. Moreover, we also show that variance reduction methods are not always the best and they may stick in local minimums when we are in a non-convex regime. We also propose a solution using momentum to overcome this problem. In the end, we present some experiments of the discussed methods when using a neural network for prediction.
View BlogSGD is the most popular algorithm for optimization in Deep Learning and Machine Learning due to its computation efficiency. In this blog, I provide the convergence proof of this algorithm with different assumptions.
View Blogدر این وبلاگ قصد دارم راجع به بدترین تجربه و انتخاب زندگیم که انتخاب مهندسی کامپیوتر صنعتی اصفهان باشه صحبت کنم. قصد دارم که خیلی جزیی و موردی ثابت کنم که چرا این دانشگاه و به خصوص دانشکده کامپیوتر یک فاجعه ی به معنای واقعیه. چیز هایی که من مینویسم راجع به مهندسی کامپیوتر این دانشگاه هست اکثرا که مطمعنم خیلیاش رو میشه به بقیه ی رشته ها تعمیم داد. ولی خب کامپیوتر صنعتی با اختلاف بدتر از بقیه ی رشته هاست. و شما ممکنه در بعضی موارد نظر مخالفی نسبت به من داشته باشید اگه رشتتون کامپیوتر نبوده.
View Blog