About

Neural networks learn remarkably well, yet the mathematics explaining why is still being written. I am drawn to that gap. I study how gradients flow, why training stays stable or falls apart, and what lets a model generalize beyond its data, and I like to build new methods from first principles.

My independent research examines logarithmic scaling applied consistently across a whole training system, from activation functions to optimizers. Together with one collaborator, I developed LogLU, a logarithmic activation function, and ZenGrad, a logarithmically formulated optimizer. We tested them against baselines on classification, segmentation, and language model tasks, released them as open-source packages for PyTorch and TensorFlow, and are preparing the work for CVPR 2027.

I am a first-year student in the Master's in Artificial Intelligence (MAI) program at Penn State Great Valley, where I also research vision-language-action models with persistent memory. I see machine learning, deep learning, and large language models as one interconnected field rather than separate topics. Ideas from optimization shape how language models train, and insights from vision and multimodal learning feed back into how we think about learning itself. My interests lie in every part of that picture.

I serve as a reviewer for ICLR and NeurIPS. I am currently seeking research internship opportunities in machine learning and AI, and I welcome conversations and collaboration on industry research projects.

Research

Gradients through Logarithmic Scaling Formulations

Poorni Murumuri and Rishi Chaitanya Sri Prasad Nalluri. Manuscript in preparation for CVPR 2027.

We introduce LogLU and the ZenGrad and M-ZenGrad optimizers, a family of logarithmic formulations with convergence guarantees. The methods improve gradient flow and training stability across vision and language benchmarks, supported by extensive ablation and hyperparameter studies. On ImageNet-1K, validation loss falls by 54% relative to the baseline.

In-memory vision-language-action modeling

Ongoing, with Dr. Thao Minh Le and Dr. Chengfei Wang

This project augments a vision-language-action framework with a persistent memory module. The aim is to retain long-horizon context across visual and language inputs, supporting consistent multi-step reasoning and grounded action generation in embodied agents.

Open-source software

LogLU

A logarithmic activation function that improves gradient behavior and training stability. Available for PyTorch and TensorFlow.

pip install loglu-pytorch   pip install loglu-tensorflow

PyTorch packageTensorFlow package

ZenGrad

A gradient-based optimizer with logarithmic learning-rate scaling, designed for stable and efficient convergence. Available for PyTorch and TensorFlow.

pip install zengrad-pytorch   pip install zengrad-tensorflow

PyTorch packageTensorFlow package

Patents

Indian patent applications

  • GrantedPerformance evaluation of an activation function for training deep neural networks. Application 202341087256. View
  • GrantedSystem and method for prediction of bio-oil production via hydrothermal liquefaction. Application 202241056535. View
  • PublishedComputer-implemented system for predicting drug-target interactions. Application 202541054965. View
  • PublishedSystem and method for building a forecasting model for biogas production. Application 202341074196. View

Experience

Graduate researcher, Pennsylvania State University

September 2026 to present

Research on memory-augmented vision-language-action models for embodied agents.

Independent researcher

May 2026 to present

Developed LogLU, ZenGrad, and M-ZenGrad, and released them as open-source packages.

Peer reviewer, ICLR 2025, ICLR 2026, NeurIPS 2026

September 2024 to present

Reviewed about ten submissions in machine learning, deep learning, and large language models, providing detailed technical feedback to authors.

Undergraduate researcher, SRM University AP

September 2021 to March 2025. Supervisors: Dr. Karthik Rajendran, Dr. Prabakaran G, Dr. Ashu Abdul, Dr. Ajay Dilip Kumar

Led applied machine learning projects in bio-energy, finance, and drug discovery. Work included live stock-market forecasting with ARIMA models deployed on TradingView, biogas production forecasting with ARIMA and SARIMAX (93% accuracy, MAPE under 7%), and drug-target interaction modeling on the BindingDB and Davis datasets.

Education

Pennsylvania State University, Great Valley. M.S. in Artificial Intelligence, August 2026 to December 2027 (expected).

SRM University AP, Amaravati. B.Tech in Computer Science and Engineering (AI and ML), August 2021 to October 2025. CGPA 8.43/10.

Conference presentations

  • Feb 2024Oral presentation, 16th International Conference on Machine Learning and Computing (ICMLC), Shenzhen, China. Certificate
  • Mar 2024Poster, AESEE 2024, SRM University AP, on machine-learning prediction of bio-oil yield. Certificate
  • Dec 2022Poster and flash talk, BSBB 2022, IIT Guwahati, on machine learning for bio-oil prediction. Certificate

Honors and awards

  • 2024Travel grant of INR 77,000 to present at ICMLC 2024, SRM University AP
  • 2023Silver Medal, SRM Research Day, 7th edition, for the LogLU activation function. Certificate
  • 2023Silver Medal, SRM Research Day, 6th edition, AI and ML track. Certificate
  • 2022Travel grant of INR 22,000 to present at BSBB 2022, SRM University AP
  • 2022Supervised Machine Learning: Regression and Classification, Stanford University. Certificate
  • 2022First runner-up, WEBMOBDEVTHON hackathon. Certificate
  • 2021Winner, SMACATHON, for a smart-city transportation concept. Certificate
  • 2021Participant, PRODEVTHON hackathon, building a payments and utilities app. Certificate

Contact