Karthik Ragunath Ananda Kumar

Artificial Intelligence Engineer - Machine Learning Intern, Samsung Research America (SRA) - SPEEX Labs, IMEX, MPS (Mountain View, California, USA)

• Researched and implemented deep learning architectures for neural texture rendering and 3D head mesh fitting to generate hyper-realistic 3D human avatar talking heads. This research involved implementing deep learning models to parse facial information from 2D images, predicting 3DMM coefficients, and projecting 2D images into 3D space using the predicted 3DMM-coefficients. As part of this research, I also implemented a few novel ideas that involve blending UV texture maps generated from multiple 2D image sources to improve neural texture rendering on 3D (talking) head meshes.
• Designed and implemented novel lightweight multimodal auto-regressive conditional variational auto-encoder and conditional auto-encoder deep learning architectures which are capable of generating lifelike 3D human body motions. These novel architectures were driven by input audio-context and auto-regressive pose estimations to generate smooth human-like body motions in terms of rotation angles and translation coordinate outputs for multiple joints in a real-time manner (on-device). The performances of these models were comparable to that of the heavier non-real-time SOTA models. This work of mine is currently in the evaluation phase in Samsung Research America to be published as a research paper in upcoming conferences.
• Currently exploring deep learning architectures for Non-Rigid-ICP mesh registrations and generating volumetric 3D avatar talking head meshes from scratch using multiview 2D images (with camera calibrations).

Machine Learning Intern, Samsung Research America (SRA) - NEON Team, STAR Labs (Mountain View, California, USA)

• Researching on facial-action recognition models and body-motion recognition models to recognize different facial expressions and human body movements in video frames. As part of this research, experimented on unsupervised clustering deep neural networks, temporal convolution networks, transformer architectures and few shot-learning approaches (Prototypical Networks and MAML). Achieved improvement on SOTA classification metrics in unsupervised deep clustering by introducing novelties in grouping algorithm.
• Working on Motion Variational Auto-Encoders (MVAEs) to generate realistic controllable human action videos.
• Built a PyTorch distributed training framework from scratch to make NEON’s deep-learning models to be trainable in a distributed manner. Integrated this framework into NEON's production system.
• Working on feature extractors- {Contour, Keypoint, Crop} extractors and auto audio-annotator frameworks. Additionally, built front-end and backend systems to trigger the extractor and annotation pipeline processes and also track their status.
• Performed preliminary literature survey and proposed a novel approach to sample uniform distribution across different labels in an unsupervised manner. (Yet to conduct experiments and validate).

Data Scientist - NLP and CV Research Intern, Ancestry.com (Lehi City, Utah, USA)

• Researched on "Joint Entity and Relationship Extraction" neural network models to extract relationship between entities in multi-page documents like books, blog posts, etc and convert them into interpretable graphs.
• As part of this research work, I worked on ensembling deep learning models from multiple SOTA research papers to solve the problems of:
  1. (i) capturing long range entity relationship dependencies.
  2. (ii) subject entity overlap
  3. (iii) entity pair overlap problems
  4. (iv) coreference resolution in extracting entity relationships, etc.
• Got selected to present my summer intern work in-front of executive board at Ancestry.

Software Engineer - Machine Learning/Deep Learning and Search, Mad Street Den Inc

Deep Learning and Machine Learning

• Designed and implemented Parts of Speech (PoS), Named Entity Recognition (NER), Dependency Parser Neural Network Models; LSTM based Auto-Correction and Auto-Suggestion models to optimize text-search system at production scale.
• Designed a document searching system which involved working with transformer models to generate search embeddings.
• Designed a text classifier framework to classify e-commerce products using XGBoost, MLP and SVM models.
• Experimented with XGBoost’s Learning to Rank Algorithms for fine tuning personalisation models.
• Designed recommendation algorithms using Apache Solr's boost and filter functions based on buy and search histories.

Distributed Systems Engineering

• Core member of the team which built and tuned Apache Solr (and Zookeeper) Clusters for enabling searching and listing operations on e-commerce websites.
• Scaled Apache Solr’s Distributed Computing Infrastructure to handle up to 40,000 requests per minute.
• Worked on building data pipelines for powering real time search requests using AWS SNS, Redis Queues, etc.
• Worked on setting up and tuning of Spark clusters optimizing (Py)Spark queries to optimize data digestion pipelines.
• Performed database backup, synchronization and migration operations.

Member Technical Staff (Software Engineer), Zoho Corporation

• Worked extensively in building libraries using Apple’s Core Data framework for powering offline iOS applications.
• Worked on implementing OCR algorithms on receipt images to extract information from them in the Zoho Expense iOS application.
• Built frameworks for handling image data and rendering custom animations in iOS applications.
• Completely redesigned an iOS application called "Zoho Invoice Generator" and released it in the App Store.

Thesis Research - Perception and Motion Planning Deep Reinforcement Learning

Pattern Discovery and Machine Learning (PDML) Lab, University of Texas at Dallas

As part of my on-going thesis research under the guidance of Prof. Feng Chen on Perception and Motion Planning components of Autonomous Driving Systems in the Pattern Discovery and Machine Learning (PDML) Laboratory at UT Dallas, I am experimenting with Deep Reinforcement Learning Algorithms, namely, Proximal Policy Optimization (PPO) and Soft Actor-Critic algorithms (SAC) powered by agent’s camera POV image features and also Uncertainty-Aware Object Detection components. The Uncertainty-Aware Object Detection is particularly designed to compute Out-of-Distribution (OOD) probabilities of the objects identified by the object detection component in order to augment the motion planning deep reinforcement learning algorithms. The deep reinforcement learning algorithms were trained on CARLA Autonomous Driving Simulator Engine. The simulator is wrapped as a custom OpenAI gym environment so that data streamed from the simulator engine can be accessed in real time (with synchronization) when used for training with Deep RL algorithms with the objective to plan the motion path of the agent (vehicle).

• https://github.com/Karthik-Ragunath/carla_sac {Work-In-Progress}
• https://github.com/Karthik-Ragunath/Semantic-Object-Detection {Work-In-Progress}

***Complete research code is yet to be exposed to a public Github-Repository since we are planning for a paper submission in the near future.***

Exploring Usage Of Graph Neural Networks For Encoder Networks

University of Texas at Dallas

Researched on exploring the usage of Graphical Neural Network (GNNs) variants for designing a novel encoder architecture (as part of an encoder-decoder architecture) to create abstract summaries utilizing “Joint Entity and Relationship Extraction” based approaches at UT Dallas.

• https://github.com/Karthik-Ragunath/pytorch_multi_head_selection_re {Work-In-Progress}