Skip to main content

Projects — Scalable Neural Models

I specialize in building interpretable neural networks, demonstrating that with proper representation, a model can be both high-performing and transparent. My solutions are tailored for domain experts, using techniques like domain-guided losses and architectures to address complex tasks in machine learning, natural language processing, and computational sciences.

For my PhD, I worked on developing fair and accurate neural models for toxicity detection in natural language. I used Multi Task Learning models to learn how both shared and specific tones of toxicity vary across demographic groups. This goes beyond a “one-size-fits-all” model which most often overfits to majority groups and raises algorithmic fairness concerns.

Machine Learning & Neural Networks

  • Neural SVD Solver for Big Data

    • Two-stage neural engine as alternative to randomized SVD techniques
    • Explicit memory requirement guided by feature dimension and desired rank
    • Fully interpretable model with meaningful outputs and weights

  • Hyperspectral Unmixing for Mixture Model

    • Autoencoder structure (SCA-Net) to perform blind unmixing of mixture model
    • Achieves 1000x lower RMSE and SAD scores than reported in state of the art works
    • Low-weight network with strict interpretability in terms of model

  • Streaming Low-rank Model for Generalized Rayleigh

    • Improved model for Generalized Rayleigh using low-rank constraint for streaming big data
    • Extensions to Minimum Noise Fraction for Denoising and Linear/Kernel Discriminant Analysis
    • Achieves around 10x efficiency in time and space compared to state of the art models

Computational Sciences & Imaging

  • Hyperspectral Imaging Analysis

    • End-to-end cancer detection pipelines for biopsy tissue samples
    • Low-rank, fast and memory-efficient denoising algorithms with error bounds
    • Higher-order feature selection to weed out redundant and irrelevant features

  • Real-Time CUDA Visual Saliency

    • GPU-based real-time implementation for obtaining visual saliency
    • Psychovisual saliency models for color images

  • 3D Image Segmentation for Surgical Planning

    • Real-time 3D image segmentation assist during facial reconstruction surgery
    • Exporting 3D models for surgical planning tools and 3D printing

  • Camera motion estimation for Cryo-EM images

    • Simulated Cryo-EM images with simulated noise from Protein Database structures
    • Built 3D virus capsid geometry from 2D image stacks using bundle assignment
    • Trained customed ResNet for relative pose estimation of cameras and virus class identification

Natural Language Processing

  • Multi-Task Learning Toxicity Model

    • Conditional MTL model to learn toxicity targeted at different groups
    • Improved recall ~8% and ~15% over Independent and SoA MTL models
    • Runtime and parameter reductions by ~56% and ~72% over baseline

  • GAP for Target-Group Detection

    • Group-fairness loss function based on Accuracy Parity measure
    • Balanced group accuracy around target-group detection
    • Group disparity reduced from ~22% to ~8% with minimal accuracy drop

  • Neural Pareto Optimality for Classification and Search

    • Interpretable PINN-based Pareto hypernetwork (SUHNPF) to benchmark non-convex verifiable solutions
    • Extension to finding Pareto Front for Accuracy v.s. Fairness and Relevance v.s. Diversity tasks
    • Scalable to high dimensional neural problems to trace out an approximate Pareto trade-off front

My Submitted Works

Published and Arxiv

Grouped by Type

  • Conference

    • [C10]

      Fairness-Aware Multi-Group Target Detection in Online Discussion [pdf] Gupta, De-Arteaga, Lease. FAccT 2026.

    • [C9]

      Finding Pareto trade-offs in fair and accurate detection of toxic speech [pdf] Gupta*, Kovatchev*, Das, De-Arteaga, Lease. iConf 2025.

    • [C8]

      Same Same, But Different: Conditional Multi-Task Learning for Demographic-Specific Toxicity Detection [pdf] Gupta*, Lee*, De-Arteaga, Lease. WWW 2023.

    • [C7]

      Learning a neural Pareto manifold extractor with constraints [pdf] Gupta*, Singh*, Bollapragada, Lease. UAI 2022.

    • [C6]

      Pareto Solutions vs Dataset Optima: Concepts and Methods for Optimizing Competing Objectives with Constraints in Retrieval [pdf] Gupta*, Singh*, Das, Lease. ICTIR 2022.

    • [C5]

      A Streaming model for Generalized Rayleigh with extensions to Minimum Noise Fraction [pdf] Gupta, Bajaj. IEEE International Conference on Big Data 2019.

    • [C4]

      Correlation, Prediction and Ranking of Evaluation Metrics in Information Retrieval [pdf] (Best Student Paper Award). Gupta, Kutlu, Khetan, Lease. WWW 2023.

    • [C3]

      Efficient Clustering-based Noise Covariance Estimation for Maximum Noise Fraction [pdf] Gupta, Bajaj. NCVPRIPG, Springer 2017.

    • [C2]

      A GPU based real-time CUDA implementation for obtaining Visual Saliency [pdf] Gupta*, Agarwal*, Mukhopadhyay, Layek. ICVGIP, ACM 2014.

    • [C1]

      Psychovisual saliency in color images [pdf] Gupta*, Agarwal*, Layek, Mukhopadhyay. NCVPRIPG, IEEE 2013.

    Journal

    • [J2]

      HOFS: Higher order mutual information approximation for feature selection in R [pdf] Gajowniczek, Wu, Gupta, Bajaj. SoftwareX, Elsevier 2022.

    • [J1]

      A Fully Automated, Faster Noise Rejection Approach to Increasing the Analytical Capability of Chemical Imaging for Digital Histopathology [pdf] Gupta*, Mittal*, Balla, Bhargava, Bajaj. PloS One 2019.

    Arxiv

    • [A8]

      A Scalable Multi-Task Learning Framework for Modeling Demographic Disagreements in Annotation Tasks [pdf] Gupta, De-Arteaga, Lease. Coming soon on Arxiv

    • [A7]

      Tail-Net: Extracting Lowest Singular Triplets for Big Data Applications [pdf] Gupta*, Singh*. arXiv preprint 2021.

    • [A6]

      SCA-Net: A Self-Correcting Two-Layer Auto-encoder for Hyperspectral Unmixing [pdf] Gupta*, Singh*, Lease, Dawson. arXiv preprint 2021.

    • [A5]

      Hybrid Neural Pareto Front (HNPF): A Two-Stage Neural-Filter approach for Pareto Front Extraction [pdf] Gupta*, Singh*, Lease, Dawson. arXiv preprint 2021.

    • [A4]

      Streaming Singular Value Decomposition for Big Data Applications [pdf] Gupta*, Singh*, Lease, Dawson. arXiv preprint 2020.

    • [A3]

      Extracting Optimal Solution Manifolds using Constrained Neural Optimization [pdf] Gupta*, Singh*, Lease. arXiv preprint 2020.

    • [A2]

      Prevention is Better than Cure: Handling Basis Collapse and Transparency in Dense Networks [pdf] Gupta*, Singh*, Dawson. arXiv preprint 2020.

    • [A1]

      TIME: A Fully Convolutional Neural Network Architecture with Interpretable Kernels for Dynamic Physical Processes [pdf] Gupta*, Singh*, Lease, Dawson. arXiv preprint 2020.