Hi, I'm Yuxi Yang.
Chasing the stars through
code & cognition.

A super-resolution project developed for the Google Summer of Code 2026 application with ML4SCI, upscaling low-resolution CMS detector calorimeter jet images (64×64 → 125×125) for high-energy physics analysis.

Approach

Designed and trained an ESRGAN model on CMS calorimeter jet image data, building an end-to-end PyTorch pipeline spanning data preprocessing, adversarial training, and quantitative benchmarking. Evaluated the model against bicubic interpolation and SRCNN baselines to isolate the contribution of the generative adversarial approach.

Key Results

Achieved PSNR 76.02 dB and SSIM 0.9998 on the held-out evaluation set, substantially outperforming the classical baselines and demonstrating the viability of GAN-based super-resolution for detector image reconstruction.

A from-scratch CUDA C++ implementation of 2D image convolution, benchmarked across three progressively optimized versions to quantify the impact of GPU memory hierarchy.

Implementation

Implemented 2D convolution kernels on an NVIDIA Tesla T4, using shared memory tiling to reuse pixel data across threads and constant memory to cache filter weights. Profiled each version with nvprof to attribute speedups to specific memory-access optimizations.

Key Results

Benchmarked a CPU baseline, a naive GPU kernel, and the shared-memory-optimized GPU kernel, achieving roughly a 297× speedup over the CPU baseline. Published to GitHub as cuda-convolution.

A study of how deep convolutional networks and Vision Transformers hold up under adversarial pressure, covering attack, defense, and architectural comparison on the CIFAR-10 benchmark.

Classification & Attacks

Trained a deep CNN classifier reaching 95.16% clean test accuracy, then evaluated its robustness under three attack methods — Gaussian noise, FGSM, and PGD — across perturbation budgets ε ∈ {0.01, 0.03, 0.10}, quantifying the accuracy degradation as attack strength increased.

Defense & Architecture

Implemented PGD-based adversarial training as a defense and measured the recovery in robust accuracy. Built a Tiny Vision Transformer from scratch to compare attention-based and convolutional inductive biases under the same adversarial conditions. Full pipeline built with PyTorch on Google Colab.

An end-to-end model optimization project that took YOLOv5 from a research prototype to a deployment-ready inference engine, published at ACM GAIIS 2024.

Pipeline

Designed a complete deployment pipeline: PyTorch → ONNX → TensorRT. Exported the trained YOLOv5 model to ONNX format with dynamic batch support, then compiled to TensorRT with INT8 post-training quantization using a calibration dataset for optimal range estimation.

Key Results

Achieved a 4.17× throughput improvement (789 vs. 189 QPS) with only 0.6 mAP drop and 70% model size reduction. The quantized model maintained real-time performance on edge GPU devices, demonstrating the viability of INT8 inference for vehicle detection in autonomous driving scenarios.

Publication

Proceedings of the ACM, May 2024

A rigorous benchmarking study of classical denoising algorithms, bridging variational methods and local filtering approaches under controlled experimental conditions.

Methods Evaluated

Implemented and compared 6 algorithms: MAP-L2, ROF (Total Variation), TGV2 (Total Generalized Variation), Non-Local Means, Bilateral Filter, and a classical baseline. Each method was tested across 50 COCO images corrupted with Gaussian, salt-and-pepper, and Poisson noise at varying intensities.

Technical Depth

Developed proximal gradient and Chambolle–Pock solvers for the variational models. Performed statistical performance analysis using PSNR, SSIM, and perceptual quality metrics. The study revealed that TGV2 consistently outperformed simpler TV models by preserving smooth gradients while maintaining sharp edges.

An exploration of neural radiance fields for photorealistic 3D reconstruction, comparing implicit volumetric representations with the emerging Gaussian Splatting paradigm.

Data & Pipeline

Captured a custom 59-view dataset and processed it through COLMAP for structure-from-motion and camera pose estimation. Trained a Nerfacto model using Nerfstudio, optimizing occupancy grids and appearance embeddings for high-fidelity novel view synthesis.

Comparative Analysis

Benchmarked volumetric NeRF against 3D Gaussian Splatting across three dimensions: rendering quality (PSNR, LPIPS), inference speed (FPS at various resolutions), and memory footprint. Gaussian Splatting showed significant speed advantages while NeRF maintained superior geometric consistency in fine-detail regions.