Vaibhav Vats

MS Computer Science, University of Southern California

Areas of Interest

  • Natural Language Processing

  • Worked on: Transformers, Embeddings, BERT, Multimodals, NLU, Knowledge Graphs, NER, Information Retreival, Context, Robustness, HMM

  • Machine Learning

  • Worked on: Statistical ML, SVM, Random Forest, Recommendation Systems, Pattern Recognition, Filtering, Classification

  • Deep Learning

  • Worked on: MLP, Convolutional Neural Nets, Recurrent Neural Nets, Encoder-Decoder Models, Bi-encoder, Cross-encoders

  • Software Development

  • Worked on: Python, C/C++, Web Scraping, iOS App, SwiftUI, Android App(Java), HTML/CSS, JavaScript, Angular, NodeJS, Docker, Git, Spark

  • Data Science

  • Worked on: Data Visualization, Data Analysis, Forecasting, Prediction, Feature Extraction, Database Management

Experience

Apple

May 2022 - August 2022

Machine Learning Intern, Siri Perception

Collaborated with Siri Perception team to develop a BERT-based Classifier with Acoustic features and Confusion Networks from audio files, resulting in improved performance of Siri-based Classification for better follow-ups

Built tool to visualize embeddings of different components of neural network infrastructure for optimizing the performance of multimodal transformers with multiple modalities from audio and text

Performed machine learning feature extraction and data preparation with large-scale distributed data systems in high-computation environments using PySpark, Hadoop, and hdfs

Information Sciences Institute, USC

February 2022 - April 2022

Research Assistant, Centre of Knowledge Graphs

Worked on Scalable Zero-shot Entity Linking Mechanism on WikiData with Dense Mapping.

Generated mapping of Wikipedia entity with WikiData nodes by using BERT-based Bi-encoders, Cross-encoders, and FAISS.

LogicQuad Technologies

August 2019 - June 2022

Research Assistant - Software Development

Developed tool for blind people to dictate environment & familiar faces using image-captioning in real-time via object detection and sentence generation. Devised Encoder-Decoder model & performed one-shot learning for faces with 96.18% accuracy

Created web application for AQI Prediction of cities and devised S-ARIMA statistical model and LSTM model for predictions

Constructed an android Application in Java for text messaging with users as well as chat-bots trained using the seq2seq model with LSTM cells, each trained on different emotion datasets using Python and TensorFlow

Arbunize Digital Media

July 2018 - December 2018

Machine Learning Intern

Created Recommendation System for jobs in a team of 8, implemented Resume Parsing, Stable Matching, Collaborative Filtering, and custom NER to improve job matching accuracy from 88.12% to 94% using Python, NLTK & TensorFlow.

Developed ontologies to perform ontology similarity matching using Protégé, OWL, and NLTK for improved recommendations on jobs.

Projects

Neural Decipherment of Lost Languages

Python, PyTorch, Min-Cost Flow, Seq2seq Model, CNN, BiLSTM

Built a system to decipher archaic languages using symbols of known language. The proposed model is able to identify 81% of cognates correctly between Ugaritic (lost) and Hebrew (known) languages, generating results with a 15.13% improvement. Devised Probabilistic Weight Initialization, stacked BiLSTM seq2seq architecture, Residual Connections, and minimum-cost flow algorithm to get improved results on 4 different language pairs.

Stock Search - iOS App

Swift, SwiftUI, HighCharts, FinnHub

Dynamic iOS application to trade stocks with live updates, watchlists, recommendations, user portfolios, and news. Implemented live interactable graphs, buy-sell features, card views, Finnhub and HighCharts API, and customizable layout.

DigiVision

Python, Django, MongoDB, OpenCV, Tensorflow, GRU, MTCNN

A tool that describes the environment & recognizes humans for visually-impaired people. It uses Object Detection Model as Encoder and GRU model for Sentence Generation as Decoder. Implemented ImageNet for Human Detection and One-Shot learning for face verification. Technologies used are TensorFlow, Convolutional Neural Networks & Recurrent Neural Networks.

Stock Search - Web

Angular, NodeJS, Express.js, Bootstrap, HTML/CSS

Built a responsive web application that allows users to search and trade stocks. The application has a watchlist and separate dashboard for monitoring stock trade with hourly change charts, profit and loss tracking, recent news, and recommendations. The front end is built using Angular10, AJAX, and Bootstrap. The server is built using NodeJS.

Zephyr - Breathe Clean

Python, JavaScript, HTML, CSS, Django, Scrapy, Tensorflow, LSTM

Site for the prediction of AQI for 39 cities along with Articles, discussions and graphical representation of statistics. Used Statistical Model like S-ARIMA for long-term prediction and Deep Learning for Daily Predictions. Use of TensorFlow, Recurrent Neural Networks, Django, Scrapy & data collected manually alongside APIs.

Checkers Game Playing Agent

Python, Minimax Algorithm, Alpha-Beta Pruning for better efficiency

Created Game Playing Agent for Checkers in Python that makes a valid move according to the rules of Checkers. It uses Minimax Algorithm for deciding which is a better move. Further implemented Alpha-Beta Pruning to Minimax Algorithm to increase the efficiency at greater depths by 70%.

FOL Resolution

Python

Developed a system that takes First Order Logic queries as Input and performs a Full Resolution on it. The system outputs a logical conclusion with 100% accuracy.

Sorting Visualizer

Python, Plotly, Dash, JavaScript, HTML, CSS

A sorting visualization tool that animates various types of sorting algorithms like Merge Sort, Heap Sort, Selection Sort, etc. Animations are carried out using Plotly for Python and Chart.js for Web. JavaScript and Dash are used for Web. Jupyter notebook is used to run visualizations offline.

RelaxBot

Java, Python, Firebase, Tensorflow, NLP

An Native Android application for communication among friends. Consists of Chatbots trained using Deep Learning as add-ons. Built using Firebase, RNN, Seq2seq model, tensorflow, keras, and other standard Android tools and APIs. Languages used are Java and Python.

Drawing Detection

Python, OpenCV, SVM, HOG features, Scikit

A tool that recognizes numbers drawn in air using a particular colored object in real-time. It tracks the object, store its path and converts the formed image into Histogram of Gradient (HOG) form. Then uses SVM to classify the shape to the nearest matching result.

Art Generation

Python, Matplotlib, Neural Style Transfer, VGG net

It creates art by merging two images, namely, a content image (C) and a style image (S), to create a generated image (G). The generated image G combines the content of the image C with the style of image S. It is similar to effects used in apps like Prisma.

Shust-It

Java, Firebase, Google Location API

An Android app which triggers silent mode during a certain period of time or if the user enters a certain region. It is developed in native Android environment using Java with core Classes along with Google APIs.