Portrait of Soroush Mesforush

Soroush Mesforush

PhD Student at UCSD

Projects:

Universality in high-dimensional regression

Illustrative schematic connecting structured covariates to sparse estimation
Illustrative schematic of the universality question; this is not an empirical result.

We investigate universality in the high-dimensional sparse regime with linear dependency. That is, when do certain properties for sparse estimators continue to hold under structured, linearly dependent covariates.

  • Analyzing Lasso in sparse regimes with dependent covariate models.
  • Investigating universality in high-dimensional regression and its implications for compressed sensing.
  • Developing theoretical and simulation frameworks for sensing matrices with spatial and temporal dependence.

Related manuscript: Lasso Universality Under Linearly Dependent Covariates in the Sparse Regime .

Compressive sensing for massive MIMO detection

Bit error rate comparison for a 32 by 32 BPSK massive MIMO system, showing the proposed detector close to the benchmark
BER comparison for BPSK with K = N = 32 (Figure 1 of the ICEE 2024 paper).
Runtime comparison between the benchmark and proposed detector as the number of receivers increases
Runtime comparison for BPSK as the number of receivers increases (Figure 3 of the ICEE 2024 paper).

My undergraduate thesis at the University of Tehran developed a sparsity-based detector for uplink massive MIMO systems. The project formulated detection through sparse representation and implemented an iterative scheme designed for efficient runtime and low computational complexity.

Publication: Efficient Signal Detection via Compressive Sensing in Uplink Massive MIMO Systems.

Learning-based perception and generative models

Object-detection output with bounding boxes around pedestrians and vehicles (image quality has been enhanced)
Object-detection output from an earlier intelligent-systems project.

During a summer internship at HARA AI, I studied neural-network fundamentals and applied deep-learning methods to computer-vision tasks, including extending YOLOv7 to support rotated bounding boxes. My later coursework explored diffusion models, transformers, variational autoencoders, and generative adversarial networks.

  • Projected diffusion planning for kinodynamic motion planning with optimization-based safety projection.
  • Prompt-conditioned multimodal story and image generation.
  • VAE-based chest X-ray generation and DCGAN-based WikiArt synthesis.