Sampras Dsouza

Hi I am Sampras. Currently, I am MSCS Student at NYU courant school of Mathematics. Previously, I have worked as a Software Enginer 2 (4+ Years of Experience) at Cimpress India. I graduated from Dwarkadas J. Sanghvi College of Engineering with a major in Information Technology with (9.47/10 CGPA). and worked as a research intern under prof. Jatin Batra.

Google Scholar  /  Email  /  CV  /  Github  /  Linkedin

Research

My research interest broadly lies in Computer Vision, Deep Learning and Cognitive Science.

News

🎉 10-2026 Our paper Scalable Fully Bayesian Gaussian Process Classification for Cancer Histopathology: Calibrated Uncertainty Under Distribution Shift was accepted at the NeurIPS 2026 ASCI Workshop!

Projects

• Developed a fully Bayesian Gaussian Process Classification model for cancer diagnosis using accelerated Randomly Pivoted Cholesky (RPCholesky) and Hamiltonian Monte Carlo (HMC) to enable scalable uncertainty-aware classification.
• Applied the method to large-scale histopathology datasets including PCam and CAMELYON17-WILDS, using Phikon-based 512-dimensional embeddings extracted from pathology image patches.
• Reduced the computational cost of Gaussian Process inference by approximating the dense N × N kernel matrix with a low-rank factorization, making Bayesian GP classification tractable for datasets with nearly 300K training examples.
• Implemented non-centered HMC posterior sampling and a pivots-as-inducing-points predictive scheme to improve scalability while preserving calibrated Bayesian uncertainty estimates.
• Compared the proposed Low-rank HMC GP against SVGP and neural network baselines such as U-Net, evaluating performance using accuracy, AUROC, ECE, NLL, and Brier score.
• Achieved substantially better calibration under distribution shift on CAMELYON17-WILDS, reducing ECE by 87.3%, NLL by 44.5%, and Brier score by 76.8% compared to the SVGP baseline.

Project Report  /  Demo Video  /  Code

• Pretrained VICReg-based self-supervised models on a 700K-image custom dataset and evaluated representations on downstream image classification tasks under ImageNet-style constraints.
• Trained ResNet-50×2 architectures (<100M parameters) using large-batch training (batch size 1024) with the LARS optimizer for stable non-contrastive SSL.
• Optimized data augmentations, loss coefficients (variance, invariance, covariance), and hyperparameters to prevent representation collapse and ensure convergence.
• Performed linear probing and finetuning on downstream tasks, analyzing the impact of architecture and augmentation choices on transfer performance.

Project Report  /  Demo Video  /  Code

Publications

The main aim of our project is to develop a system that assists a visually impaired person in all ways possible. Existing aids consist mainly of guard dogs which are very costly to purchase and canes that are not feasible to be used in all possible conditions(Eg: Outdoor Navigation). We make use of the computing power of the raspberry pi to develop a portable system that assists a visually impaired person in all possible situations. Our system is implemented on a Raspberry Pi 3 model having a single Pi camera unit. The system consists of three main modules, first is the navigation module that is equipped with a depth-sensing technique that can be used by the user for outdoor navigation. Second, is the face and object detection module that can be used by the user to recognize various objects and faces in front of him. Third, is the text recognition facility can be used by the user to read the text. All the communication between the system and the user takes place with the help of the microphone that is connected to the 3.5mm audio port of the raspberry pi.

Selected Awards and Honors

06-2024: I received the NPTEL Distributed System Certification with 80% Link to Certification.

Service

Reviewer: TNNLS