Sarath Shekkizhar, Ph.D.

San Francisco Bay Area, California shekkizh.github.io shekkizh@usc.edu shekkizh shekkizh

Staff (Data) Scientist, Salesforce October 2024 - Present

Education

Ph.D. in Electrical and Computer Engineering Aug 2017 - May 2023

GPA: 3.93

University of Southern California, Los Angeles, CA
Advisor: Antonio Ortega.

M.S. in Computer Science Aug 2017 - May 2022

GPA: 4.0

University of Southern California, Los Angeles, CA

M.S. in Electrical Engineering (Computer Vision, Machine Learning) Aug 2012 - Dec 2013

GPA: 3.86

University of Southern California, Los Angeles, CA

B.Tech. in Electronics and Communication July 2008 - June 2012

GPA: 9.12

National Institute of Technology, Tiruchirappalli, India

Publications

Reasoning in Large Language Models: A Geometric Perspective

R Cosentino, S Shekkizhar, arXiv Preprints, 2024

Characterizing Large Language Model Geometry Solves Toxicity Detection and Generation

R Balestriero, R Cosentino, S Shekkizhar, International Conference on Machine Learning (ICML), 2024

Towards a geometric understanding of Spatio Temporal Graph Convolution Networks

P. Das, S. Shekkizhar, A. Ortega, IEEE Open Journal of Signal Processing, 2024

Data Sampling using Locality Sensitive Hashing for Large Scale Graph Learning

S. Shekkizhar, N. Bulut, M. Farghal, S. Tavakkol, M. Bateni, A. Nandi, International Workshop on Mining and Learning with Graphs, Conference on Knowledge Discovery and Data Mining (KDD), 2023

A data-driven graph framework for geometric understanding of deep learning

S. Shekkizhar, A. Ortega, Graph Signal Processing Workshop 2023, 2023

Study of Manifold Geometry using Multiscale Non-Negative Kernel Graphs

C. Hurtado, S. Shekkizhar, J. Ruiz-Hidalgo, A. Ortega, IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022

The geometry of self-supervised learning models and its impact on Transfer learning

R. Cosentino, S. Shekkizhar, M. Soltanolkotabi, S. Avestimehr, A. Ortega, arXiv Preprints, 2022

NNK-Means: Data summarization using dictionary learning with non-negative kernel regression

S. Shekkizhar, A. Ortega, IEEE 30th European Signal Processing Conference (EUSIPCO), 2022

Channel redundancy and overlap in convolutional neural networks with Channel-wise NNK graphs

D. Bonnet, A. Ortega, J.Ruiz-Hidalgo, S.Shekkizhar, IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022

Channel-Wise Early Stopping without a ValidationSet via NNK Polytope Interpolation

D. Bonnet, A. Ortega, J.Ruiz-Hidalgo, S.Shekkizhar, Asia Pacific Signal and Information Processing Association (APSIPA), 2021

Model selection and explainability in neural networks using a polytope interpolation framework

S. Shekkizhar, A. Ortega, *Asilomar Conference on Signals, Systems, and Computers *, 2021

Revisiting local neighborhood methods in machine learning

S. Shekkizhar, A. Ortega, IEEE Data Science and Learning Workshop (DSLW), 2021

Efficient graph construction for image representation [Best student paper]

S. Shekkizhar, A. Ortega, IEEE International Conference on Image Processing (ICIP), 2020

Graph-based Deep Learning Analysis and Instance Selection

K. Nonaka, S. Shekkizhar, A. Ortega, IEEE International Workshop on Multimedia Signal Processing (MMSP), 2020

DeepNNK: Explaining deep models and their generalization using polytope interpolation

S. Shekkizhar, A. Ortega, arXiv Preprints, 2020

Graph Construction from Data by Non-Negative Kernel Regression

S. Shekkizhar, A. Ortega, IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2020

Neighborhood and Graph Constructions using Non-Negative Kernel Regression

S. Shekkizhar, A. Ortega, *Under review (IEEE Transactions on Pattern Analysis and Machine Intelligence) *, 2019

M. Plihal, E. Soltanmohammadi, S. Paramasivam, S. Ravu, A. Jain, S. Shekkizhar, P. Uppaluri, US Patent Office, 2019

Detection and removal of Salt and Pepper noise in images by improved median filter

S. Deivalakshmi, S. Shekkizhar, P. Palanisamy, IEEE Recent Advances in Intelligent Computational Systems, 2011

Work Experience

Member of Technical Staff, Tenyx (Acquired by Salesforce) June 2023 - October 2024

Part of the founding team at Tenyx building Voice AI for customer support. Was primarily focused on research and creating IP. Key accomplishments include research on continual learning [VB article], building TenyxChat series of models [VB article], and geometric characterization of LLMs [HN Discussion]. Was also involved in product development, particularly in endpointing, audio disambiguation, and agent governance.

Research Intern, Google, Sunnyvale, CA Sep 2022 - Dec 2022

Host: Mohamed Farghal, Animesh Nandi, Behavior Protections, Counter-Abuse Technology.
Worked on understanding the impact of input data used in training graph models and scalable sampling approaches to improve semi-supervised graph learning. Preliminary experiments with proposed graph learning showed 3x increased recall in abuse detection.

Software Engineer 2, KLA Tencor, Milpitas, CA Mar 2014 - Oct 2016

Designed and developed tools to classify and visualize defect modulations for Process Window Qualification in wafer fabrication. Also, implemented and co-owned components for analysis and classification using decision trees and random forests.

Freelance Researcher, Toonchat (Demo: youtu.be/B7LyoWksHHE) Jan 2014 - Jun 2014

Researched and worked with animators and researchers on real time face tracking under the advise of Dr. Eric Petajan for low bandwidth animations and anonymous video chat clients

Software Developer, Laboratory of Neurological Imaging, USC (invizian.loni.usc.edu) Aug 2013 - Dec 2013

Worked under the supervision of Dr. John Van Horn as part of the Information Visualization project, a platform to interact and research on large amounts of brain data

Other Research Works

Manifold embedding using NNK Graphs Jan 2020 - May 2020

Revisited data embedding using graphs in terms of robustness and stability with respect to hyperparameters. NNK graphs are significantly sparser compared to other graph constructions, while being able to capture the structure of noisy manifolds.

Manifold Regularized Variational Autoencoder (VAE) Aug 2019 - Dec 2019

Studied disentanglement in VAEs with explicit regularization using graphs. This work was motivated from the perspective of locality often enforced in autoencoders using noisy sampling of embeddings.

Are combined Fuzzy Cognitve Maps (FCM) always better than individual maps? Aug 2018 - Dec 2018

Analysed the performance of decisions taken by individuals in a simple game against that of the additive. Combined FCM reduces the effect of error associated with each individual.

Impact of lp-norm choice on K-nearest neighbor graph construction Jan 2018 - May 2018

Explored the impact of distance norms for k-nearest neighbor graph construction in high dimensional spaces using eigen analysis. Lower norms produce data clusters better than euclidean and higher norms.

Graph based Image Segmentation, Prof. Antonio Ortega Aug 2013 - Dec 2013

Performed experiments and analysis on graph based approach to image Segmentation. Leveraged methodologies for finding approximate Fiedler vector on a graph laplacian as an alternative to doing normalized cuts.

3D Face Recognition System, Dr. Jongmoo Choi, Prof. Gerard Medioni May 2013 - Aug 2013

Developed on the core recognition library and created an evaluation framework and data set for benchmarking. Made integration efforts on incorporating 3D modelling module for recognition.

Dynamic Face Warping, Prof. Antonio Ortega Jan 2013 - June 2013

Implemented a real time facial tracking and warping module in DaVinci DSP board. The project emphasized working under constrained resources and was targeted towards applications in mobile.

Optical Character Recognition, Prof. S. Deivalakshmi Jan 2012 - June 2012

A neural network based character recognition system for use with motor vehicle license plate recognition was developed. The system was evaluated with different fonts and lighting confitions.

Classification of Mammograms, Prof. S. Deivalakshmi Aug 2011 - Dec 2011

A method to differentiate and identify the nature of tumor in mammograms using discriminant analysis on extracted features was developed and evaluated on the MIAS database.

Open-source Works

Deeplearning Projects using Tensorflow (github.com/shekkizh/TensorflowProjects)

Highlights: DCGAN for generating flowers/ faces, Face completion using context, Deep dream, VGG visualization, Image Inversion, Style Transfer

Neural Networks Experiments (github.com/shekkizh/neuralnetworks.thought-experiments)

Experiments on Activation functions, Model Pruning (Optimal Brain Damage), Unsupervised Learning using AutoEncoders, VAEs, GANs

Fully Convolutional Networks for Semantic Segmentation (github.com/shekkizh/FCN.tensorflow)

Tensorflow implementation of FCNs for segmentation as in CVPR paper applied on MIT scene parsing challenge dataset

Energy Based Generative Adversarial Networks (github.com/shekkizh/EBGAN.tensorflow)

Model implementation of Junbo et. al’s paper of training GAN with energy based objective in tensorflow

Image Colorization (github.com/shekkizh/Colorization.tensorflow)

Experiments on leveraging CNNs for colorizing grayscale images by statistical knowledge gained about objects and colors from a dataset.

Image Processing Projects (github.com/shekkizh/ImageProcessingProjects)

Highlights: Eye Tracking, Facial Tracking and Localization, Seam carving, Image Stitching, Image calibration, Image filters, Object detection and processing

Co-Mentoring

Carlos Hurtado Comín, Universitat Politècnica de Catalunya (Visiting Researcher, USC) Mar 2022 - June 2023

Aryan Gulati, University of Southern California (CURVE program, USC) Aug 2021 - June 2023

David Bonet Solé, Universitat Politècnica de Catalunya (Visiting Researcher, USC) Dec 2020 - Aug 2021

Keisuke Nonaka, KDDI Research (Visiting Researcher, USC) Aug 2019 - July 2020

Teaching Experience

Course Producer, CS 561 Foundations of Artificial Intelligence, Dr. Sheila Tejada Fall 2013

Course Grader, EE 483 Introduction to DSP, Prof. Edgar Satorius Summer 2013

Course Grader, EE 483 Introduction to DSP, Prof. Edgar Satorius Spring 2013

Academic and Co Curricular Activities