JOY LUNKAD Deep Learning Specialist Pune, India joylunkad@gmail.com EDUCATION Bachelor of Computer Science in Intelligent Systems CGPA 8.9 Maharashtra Institute of Technology Aug,2019 - 23 CERTIFICATIONS Deep Learning Specialization Deeplearning.ai Aug,2020 • Neural Networks and Deep Learning • Improving Deep Neural Networks • Structuring Machine Learning Projects • Convolutional Neural Networks • Sequence Models Natural Language Processing Specialization Deeplearning.ai Nov,2020 • Natural Language Processing with Classification and Vector Spaces • Natural Lang uage Processing with Probabilistic Models • Natural Language Processing with Sequence Models • Natural Language Processing with Attention Models TensorFlow: Advanced Techniques Specializatio n Deeplearning.ai Dec,2020 • Custom Models, Layers, and Loss Functions with TensorFlow • Custom and Distributed Training with TensorFlow • Advanced Computer Vision with TensorFlow • Generative Deep Learning with TensorFlow CS50 - Introduction to Artificial Intelligence with Python Harvard | 2020 • Built up a foundation in Graph search algorithms, Classification, Optimization, Reinforcement learning, Uncertainty and Neural Networks Introduction to Algorithms MIT OCW | 2020 • Data Structures, Sorting, Hashing, Search Algorithms, Dynamic Programming Introduction to Computational Thinking and Data Science MIT OCW | 2020 • Optimization, Simulations, Basi cs of machine learning Introduction to Computer Science and Programming in Python MIT OCW | 2019 • Helped me achieve a great foundation in python ACHIEVEMENTS & SKILLS All India Rank 27 – Technothelon Indian Institute of Technology - Guwathi • Secured AIR 27 amongst more than 220,000 students in an IQ and logic reasoning test. 2017 Played Cricket for 13+ years • Amassed more than 40 awards over the years. • Since I started pursuing cricket when I was 4, throughout my long career, I have almost always played with teammates who were much older than me. I was the captain of multiple teams across multiple years. Competitive Coding • 100+ problems solved on Project Euler, Leetcode, Hackerrank, etc. ranging from the easiest to the hardest. Black Belt Holder in Shotokan Karate Deep Learning Computer Vision NLP Tensorflow Python Jax Time Series Prediction ML Techniques Data Augmentation PERSONAL PROJECTS Replicating OpenAI's CLIP model for zero shot classification. • Dual Encoder Architecture Trained with Contrastive Loss on Visual Genome Dataset. • Achieves 85% Zero Shot Multi - Label Accuracy on PASCAL - VOC Dataset. Neural Machine Translation | Translating English to Hindi • Solved it using both: with Transformers and with LSTMs + Attention. Replicated YOLOv2 for Object Detection from scratch • Trained Using a Loss Function calculated using class loss, bounding boxes loss, and confidence loss on MS COCO dataset. Semantic Segmentation | Aerial Images Using a Drone • Semantically Segmented 20 classes using a U - Net model with only 400 Images. Text Summarizer | Chatbot • B uilt Transformers implemented using JAX and TRAX Others • Implementing Google's PageRank Algorithm. • Developed games such as Nim using reinforcement learning, Minesweeper using Propositional Logic, Tic Tac Toe using Searching Techniques and Crossword by sati sfying constraints. • Simulating a robotic vacuum cleaner. RESEARCH PROJECTS Project AiM - Memory for Ai • Invented attaching memory to neural networks that is almost like human memory. • Proved that it would significantly improve the performance for convolutional neural networks (CNNs). • Achieved it using relatively cheap operations. • Authored in TowardsAi publication Dynamic Gesture Recognition • End to End Detection and classification of Hand Gestures using Deep Learning in real time. • Replicated a couple of SOT A methods to solve dynamic gesture recognition namely, 1) Using a light weight detector and a deep classifier 2) Using Motion Fused Frames - A data level fusion strategy. Humpback Whale Identification • To identify individual humpback whales in images by their names. • Used Self - Supervised Learning and heavy data augmentations to pre - train. • Trained a Siamese network using Facenet's Triplet loss where the triplets get progressively harder in regular training intervals. EXPERIENCE Indian Meteorological Department (Govt. of India) – Research Project Lead Deep Learning Researcher May, 2021 - Present • To beat the SOTA in weather prediction, I conceptualized a novel end - to - end deep learning architecture for Govt of India • The architecture was created by using various components from SOTAs of Object Detection, Time Series Prediction and others. • Proof of concept is bei ng funded by MIT and IMD. Indian Council of Agricultural Research (Govt. of India) – Research Project Lead Machine Learning Researcher Nov, 2020 - Present • Developed many different models for solving challenging problems in viticulture such as leaf and shoot counting, diseases , and insect detection, bud classification and counting, growth phase detection, etc.