Krishna Somandepalli

PhD Candidate, Electrical & Computer Engineering

I build robust machine intelligence models by learning from multiple views and modalities in datasets. A few application areas of my research are understanding multimedia content, computational imaging, affective computing and developing inclusive technologies. I enjoy working on open-ended problems that contribute to products with a tangible societal impact.

Location

Venice, California, United States

Email

somandep@usc.edu

Scholar

Google Scholar

Website

sail.usc.edu/~somandep

Experience

Aug 2015 – present

Research Assistant at SAIL, Viterbi School of Engineering, USC

Highlights

• Led research and collaboration of the Multimedia Intelligence and Content Analysis (MICA) group at SAIL with 10 PhD students
• Created three large-scale open-source datasets for computer vision and speech understanding of multimedia content
• Co-created a patent-pending application to analyze language and characters in screenplays and piloted the tool for over 25 film/TV scripts of 5 production companies including Disney and NBCUniversal
• Collaborated with Geena Davis Institute on the Project SeeJane to computationally analyze over 600 movies from 2014-2019
• Partnered with J Walter Thompson to analyze portrayal of women in over 9000 ads nominated to Cannes Lions from 2006-2017
• Co-developed an application for tracking female participation for the United Nations ITU conferences analyzing over 120 hours of meetings

May 2020 – Aug 2020

Research Intern at Google LLC.

Highlights

• Demonstrated the use of federated learning (FL) for classifying 42 emotion labels at scale in a privacy-preserving manner
• Benchmarked emotion experience and perception federated models on a dataset of ~1500 users and ~3200 raters
• Proposed an algorithm to model raters and estimate their reliability in a distributed setting with a relative improvement of 3% average precision over the state-of-the-art

May 2019 – Aug 2019

Research Intern at Google LLC.

Highlights

• Developed self-supervised methods to learn robust features of facial movements and audio from ~3M unlabeled videos
• Improved emotion classification system performance by a relative gain of 5% average precision using self-supervised features

May 2018 – Aug 2018

Research Software Intern at Google LLC.

Highlights

• Designed a data collection framework to localize cartoon characters in animated content
• Adopted computer vision models to detect, track & identify characters in video content and evaluated the system for seven animation genres

Oct 2012 – Aug 2015

Jr. Research Scientist at NYU Langone Medical Center

Highlights

• Automated systems for quality control and denoising of functional MRI data reducing data attrition by ~12% over a 6 month period
• Contributed to 5+ journal publications in the field of computational neuroimaging
• Studied test-retest reliability of the fMRI connectome and replicability of the measures in a sample of 120 participants

Education

Aug 2015 – present

PhD in Electrical & Computer Engineering from University of Southern California

Courses

• Advisor: Prof. Shri Narayanan
• Thesis: Learning Shared Subspaces across Multiple Views and Modalities

Aug 2011 – Sep 2012

M.S. in Electrical & Computer Engg from University of California at Santa Barbara

Courses

• Advisor: Prof. Matthew Turk

Aug 2007 – Jun 2011

B.E. in Electronics & Communication Engg from University Visvesvaraya College of Engineering

Courses

• Advisor: Prof. Deepa Shenoy

Publications

I have published in top peer-reviewed journals such as Transactions on Multimedia, Science Advances and conferences such as EMNLP and AAAI. For the full list of my publications, see here.

Invited Talks

Apr 2018 | Global Symposium on Gender in Media, Los Angeles - link

Automating character representation and gender attribute analysis in Hollywood movies; Attendees: 300

Sep 2017 | Global Symposium on Gender in Media, New York - link

Understanding gender portrayals in movies using multimodal analysis; Attendees: 250

Skills

Programming

Level: Master

Keywords:

• Python
• C++
• R
• TensorFlow

Level: 2/3

Keywords:

• HTML
• Javascript
• Julia
• PyTorch

Concepts

Level: Master

Keywords:

• Statistics
• Machine (Deep) Learning
• Data Analysis
• Optimization

Soft skills

Level: Master

Keywords:

• Project leadership and management
• Critical thinking
• Communication

Interests

Research Areas

Keywords:

• Multiview / Multimodal signal representatiom learning
• Unsupervised / Self-supervised learning
• Explainability in Neural Networks
• Machine learning robustness and fairness

Application domains

Keywords:

• Computational Media Intelligence
• Computational Imaging
• Computational Social Sciences
• Building inclusive technologies

Service

Peer-review

Fluency: Reviewed 5+ journal articles and 10+ conference papers in machine learning/signal processing.

Organizing

Fluency: Organizer, Workshop on Media Analytics for Societal Trends : 2018-2020;
MAST-I
MAST-II
MAST-III

Awards

Fluency: Received an award for Outstanding contribution in revieweing from Elsevier
Phase 1 Finalist in the 2021 OpenCV AI Competition from the USC team: certificate