Alireza Zaeemzadeh

PhD, Research Associate, Center for Sleep and Consciousness, University of Wisconsin, Madison.

I am a member of the center for sleep and consciousness at University of Wisconsin-Madison, where I work on integrated information theory (IIT). IIT provides a framework to analyze the neural substrates of consciousness. As a PhD student, I was a member of the center for research in computer vision (CRCV), as well as communications and wireless networking lab (CWNLab), at University of Central Florida. For Summer 2020, I joined Adobe, as an Applied Research Scientist Intern. My research lies at the intersection of Machine Learning, Signal Processing, and Information Theory.


Norm-Preservation: Why Residual Networks Can Become Extremely Deep?

IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI) 2020

We prove that the skip connections in the residual blocks facilitate preserving the norm of the gradient and lead to stable back-propagation. Can we push for extra norm-preservation? We answer this question by proposing an efficient method to regularize the singular values of the convolution operator and making the ResNet’s transition layers norm-preserving.

Iterative Projection and Matching: Finding Structure-preserving Representatives and Its Application to Computer Vision

CVPR 2019

In this project, we developed a linear-time data selection algorithm and showed its effectiveness in different tasks, including active learning on UCF101 human action recognition video dataset.

Out-of-Distribution Detection Using Union of 1-Dimensional Subspaces

CVPR 2021

In this project, we claim that we can improve the out-of-distribution detection performance by constraining the representation of in-distribution samples in the feature space. Particularly, if we embed the training samples such that the feature vectors belonging to each known class lie on a 1-dimensional subspace, OOD samples can be detected more robustly with higher probability, compared to a class-conditional non-degenerate Gaussian embeddings.

Face Image Retrieval with Attribute Manipulation

ICCV 2021

In this project, we developed a face image rertieval framework that finds the most similar faces to a query face, while giving the user the ability to both adjust and assign importance to any arbitrary subset of attributes. Work done as part of an internship at Adobe Inc. Patent pending.

A Bayesian Approach for Asynchronous Parallel Sparse Recovery

Asilomar 2018

We propose an asynchronous parallel sparse recovery algorithm which incorporates a probabilistic framework that assigns reliability scores to each processor.

piecewise constant nonnegative matrix factorization


We propose a nonegative matrix factorization (NMF) algorithm that can factorize the data into its components with piecewise constant coefficients.