Hi, I'm Aniruddha Kalkar.
Striving towards improving accessibility by building scalable solutions in the vision-language domain using Deep Learning, Computer Vision and Natural Language Processing.
About
I am a Computer Science (Artficial Intelligence) Gradute Student at University Of Southern California. A seasoned machine learning engineer with research experience in Computer Vision, Natural Language Processing and Reinforcement Learning. Proficient in developing reproducible processes and robust, production level solutions. I enjoy solving interesting problems and algorithm design and analysis. Currently working as Machine Learning Engineering Intern with the Data Science Team at the Blackberry Corporation. Addtionally, I have 2+ years of Frontend Developer experience working with JavaScript, Python, Java.
Looking for an opportunity to work on challenging problems that enable me to leverage my skills in Machine Learning and Software Engineering, have interesting experiences and professional and personal growth.
Experience
- Detected and categorized malicious programs by developing machine learning models to identify threats to users.
- Built and maintained the treat analysis data sources by collaborating with the data engineering team.
- Tools: Python, Pytorch, Bitbucket, AWS, AWS Sagemaker, Prefect, Docker
- Designed and executed experiments to train 8 different Quality Diversity Algorithms with customized reward signals in 6 reinforcement learning environments like “Slime Volley” and “Car Racing”.
- Analysed effects of learning rates on the optimal score in the RL environments.
- Co-Authored “Training Diverse High-Dimensional Controllers by Scaling Covariance Matrix Adaptation MAP-Annealing”.
- Tools: Python, Pytorch, Github, Docker, OpenAI Gym
- Enhanced VR Game developed to improved skilled locomotion for individuals with neurological impairments.
- Enabled enhanced analysis by randomizing all object locations with 0\% loss of experiment repeatability.
- Introduced functionality to store additional user action data to analyze user responses at multiple granularities.
- Tools: C#, Unity, Virtual Reality
- Added features to the METRANS student website to increase student engagement on the website by 10%.
- Developed 3 game scenes for a game using Unity to educate kids about public transportation in LA County.
- Tools: Javascript, Unity
- Optimized generative pre-trained (GPT-NEO) NLP model to auto-generate Natural language content for academic research proposals.
- Improved sentence acceptance rate by 14.7% by enhancing synonym suggestions.
- Tools: Python, PyTorch, AWS SageMaker, Generative Pre-Trained Model (GPT-Neo)
- Increased product usage across orgarnization by 63% by revamping the front end for the Lifecycle Management Service in the CI / CD Pipeline.
- Designed a prototype using NLP and Machine Learning to recommend QA testing scenarios using software requirements specification documents.
- Promoted to Software Engineering Specialist from R & D Development Associate position.
- Tools: Javascript, Python, HTML / CSS , Java, CI / CD, DevOps
- Created Novel Metric to analyze Temporal Coherence of labels placed in videos for AR Applications.
- Introduced optical flow to give upto 50x Temporal Coherence improvement for the labels placed in the videos.
- Co-authored ”SmartOverlays” published in WACV 2020.
- Tools: Python, Pytorch, Tensorflow, OpenCV
Publications
PrePrint
- We leverage efficient approximation methods in evolution strategies (ES)-based quality diversity algorithms to propose three new variants that scale to very high dimensions.
- Our experiments show that the variants outperform ES-based baselines in benchmark robotic locomotion tasks, while being comparable with state-of-the-art deep reinforcement learning-based quality diversity algorithms
- Authors: Bryon Tjanaka, Matthew C. Fontaine, Aniruddha Kalkar, Stefanos Nikolaidis
- Domains: Reinforcement Learning, Quality Diversity, Artifitial Intelligence, Machine Learning
IEEE Winter Conference on Applications of Computer Vision (WACV)
- SmartOverlays, first identifies the objects and generates corresponding labels using a YOLOv2 in a video frame; at the same time, Saliency Attention Model (SAM) learns eye fixation points that aid in predicting saliency maps for label placement; finally, computes Voronoi partitions of the video frame, choosing the centroids of objects as seed points, to place labels for satisfying the proximity constraints with the object of interest.
- In addition, our approach incorporates tracking the detected objects in a frame to facilitate temporal coherence between frames that enhances readability of labels.
- We measure the effectiveness of SmartOverlays framework using two objective metrics: (a) Label Occlusion over Saliency (LOS), and, (b) temporal jitter metric to quantify jitter in the label placement.
- Authors: Srinidhi Hegde, Jitendra Maurya, Aniruddha Kalkar, Ramya Hebbalaguppe
- Domains: Computer Vision, Machine Learning, Augmented Reality, Deep Learning, Visual Saliency
Projects
Multi-Teacher Knowledge Distillation for Visual Question Answering Systems Capstone Project
- Tools:Python, Pytorch, OpenCV
- Multi-Teacher Knowledge Distillation for Visual Question Answering Systems.
- Model size reduction up to 65x and upto 8x inference speed increase as compared to the teacher models.
HTML / UI Code Generation Web app using Flask
- Tools:Python, Flask, Keras, OpenCV, HTML, CSS3
- Constructed a template UI code generating system from input screenshots or photos of GUIs.
- Reduced code writing time by average 23.4 mins per web page.
Face sketch To Photo-Realistic Image Generation web app using Flask
- Tools: Python, Flask, Keras, OpenCV, HTML, CSS
- Spearheaded the creation of system to generate photo-realistic images from hand-drawn face sketches as well as predict age groups of people from sketches.
- Achieved 77.65 % similarity with original image and 87.38% accuracy for age group prediction.
Toxic Comment Classification Kaggle Competition.
- Tools: Python, Keras, Scikit-Learn, NLTK
- Analysis and Classification of social media comments into 6 different levels of toxicity.
- Applied Recurrent Neural Networks to classify Word embeddings from social media comments.
A Tool to detect distraction amongst drivers using video feeds and deep learning
- Tools: Python, TfLearn, OpenCV, Flask
- Designed and created a driver distraction recognition and notification program based on a live video capture.
- Attained 91.08% accuray for the 10 pre-determined distractions.
An 2 Headed - LSTM Model to predict the similarity between Quora Questions
- Tools: Python, Keras, Scikit-Learn, NLTK
- Designed and Implemented a deep learning model to predict the similarity between pairs of questions on quora.
- Used a 2 LSTM heads to generate word embeddings and and used these embeddings to calculate percentage similarity.
Skills
Languages and Databases
Python
Javascript
Java
C / C++
HTML5
CSS3
MySQL
MongoDB
Libraries
NumPy
Pandas
OpenCV
scikit-learn
matplotlib
Seaborn
Frameworks
TensorFlow
PyTorch
Keras
Django
Flask
Bootstrap
Tools And Technologies
Git
AWS
Google Cloud Platform
AWS Sagemaker
Education
University of Southern California
California, USA
Degree: Master of Science in Computer Science (Artificial Intelligence)
CGPA: 3.47/4.0
- Machine Learning
- Deep Learning and Its Applications
- Applied Natural Language Processing
- Analysis of Algorithms
- Foundations of Artificial Intelligence
- Web Technologies
Relevant Courseworks:
Walchand College of Engineering, Sangli
Sangli, India
Degree: Bachelor of Technology in Computer Science
CGPA: 8.79/10
- Data Structures
- Advanced Data Structures
- Machine Learning
- Design and Analysis of Algorithms
- Database Management Systems
- Advanced Database Management Systems
- Pervasive Computing
- Data Warehousing and Data Mining
- Statistics and Fuzzy Systems
- Distributed and Cloud Computing
Relevant Courseworks: