Amogh Dabholkar

I am Amogh Dabholkar, an MS student at Georgia Tech (Graduating in May 2023) and my focus is in the domain of Deep Learning. I completed my undergraduate degree at BITS, Pilani where I majored in Electronics and Instrumentation Engineering. I am interested in the intersection of Deep Learning and Computer Vision. More recently, I have also been exploring the new verticals like Adversarial attacks as well as Model Compression Techniques like Quantization and Pruning.

During my time at Ambarella Inc. during Summer 2022, I worked on implementing the RigL algorithm proposed by Google Brain and implementing it in TF2-Keras. I was working out of their office in Santa Clara, CA under the supervision of Sandeep Mohan and Santosh Chilkunda. At Georgia Tech, I worked with Prof Hyesoon Kim on compression of Recommendation Systems in Spring 2022. In the past, I worked at DeepEdge on building a semantic segmentation model using Mask RCNN for a Silicon Valley based client. In the preceeding summer, I interned at Vidyaroha Innovations where I helped them build an end-to-end facial recognition pipeline. Simultaneously, I was also pursuing a summer term at Carnegie Mellon University on a scholarship, where I studied graduate level Machine Learning and Image Processing courses. In my final year of undergrad, I spent six wonderful months with Prof Patricio Vela at Georgia Tech working on building a supervised learning solution for estimating Control Barrier Functions (CBFs). This work with Dr. Vela resulted in a publication at IROS 2020. Check out this seminar by Dr. Vela for more details about this work.

I am an ardent Football Club Barcelona fan and try to watch every game. I enjoy reading books a lot. Some of my favorite authors are Fredrik Backman and Sidney Sheldon. I have recently taken a liking towards Vietnamese food so hit me up with restaurant suggestions!

Feel free to check out my CV and drop me an e-mail if you want to chat with me!

~ Email | CV | Github | LinkedIn | Twitter ~

[May'23]	Graduating with MS in ECE from Georgia Tech
[Aug'22]	Returned to Georgia Tech
[May'22]	Started working at Ambarella Inc. as a Deep Learning Intern.
[Aug'21]	Started MS in ECE at Georgia Tech
[May'21]	Last day at DeepEdge as a Deep Learning Engineer
[Nov'20]	Started working with DeepEdge as a Deep Learning Engineer
[Aug'20]	Started working with DreamVu as a Computer Vision Research Fellow
[Aug'20]	Completed CMU Summer 2020 term (ECE 18-661 and ECE 18-793)
[July'20]	Graduated From BITS Pilani
[May'20]	Started CMU Summer 2020 term
[May'20]	Started working as Machine Learning Intern at Vidyaroha Innovations
[Jan'20]	Returned to BITS Pilani for Final Semester
[Aug'19]	Started working with Prof. Patricio Vela at Georgia Tech in Atlanta for my bachelor's thesis
[Aug'18]	Returned to BITS Pilani for Junior Year
[May'18]	Started working at LVPEI Center for Innovation as Summer Intern
[Aug'16]	Started undergraduate at BITS Pilani

	Deep Learning Intern \| Ambarella Inc. May '22 - August '22 Improved the Neural Network Compression toolkit by implementing RigL and adapting the implementation to work for TF2-Keras. Added key features to the algorithm which aided the accuracy recovery process after pruning.
	Special Problem Student \| Georgia Institute of Technology January '22 - present Working with Andrei Bersatti at HpArch under the Department of Electrical and Computer Engineering at Georgia Tech on compression techniques for Recommendation Systems.
	Graduate Teaching Assistant \| Georgia Institute of Technology January '22 - present Working under the supervision of Prof. Zsolt Kira as a Teaching Assistant for the on-campus version of CS7643/4644 Deep Learning for Spring 22 semester.
	Deep Learning Engineer \| DeepEdge Nov '20 - May '21 Worked on solving a semantic segmentation problem using Tensorflow2 for unconventional object classes with necessary preprocessing and refinement of predicted masks using OpenCV. Also built an auto-annotation pipeline for the company.
	Computer Vision Research Fellow \| DreamVu August '20 - October '20 Worked on deploying EfficientDet for custom object detection on panoramas captured by PAL -- world's first 360 stereo and depth sensor
	Machine Learning Intern \| Vidyaroha Innovations September '20 - June '21 Built an end-to-end Facial Recognition Pipeline using the PyTorch achieving a f1 score of 47.3 on a small dataset with picture quality on the lower end
	Controls and ML Intern \| Georgia Institute of Technology Aug '19 - Dec '19 Worked with Prof Patricio Vela at the IVA Lab in Department of Electrical and Computer Engineering at Georgia Tech. We developed a supervised machine learning approach for estimating Control Barrier Functions and the work ws published in IROS 2020.
	Summer Intern \| LVPEI Center for Innovation May '18 - July '18 Improved battery system so that it could be charged and used at the same time and to display the remaining battery. Designed the main PCB using AutoCAD Eagle.

Synthesis of Control Barrier Functions Using a Supervised Machine Learning Approach
Mohit Srinivasan, Amogh Dabholkar, Samuel Coogan, Patricio A. Vela.
IROS 2020

Cite as - M. Srinivasan, A. Dabholkar, S. Coogan and P. A. Vela, "Synthesis of Control Barrier Functions Using a Supervised Machine Learning Approach" 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2020, pp. 7139-7145, doi: 10.1109/IROS45743.2020.9341190.

Control barrier functions (CFB) are mathematical constructs used to provide formal safety guarantees.

This work presented a supervised machine learning approach for synthesis of control barrier functions, using support vector machine (SVM).

In this paper CFB is used to construct safe trajectory for robot to navigate and avoid obstacles

Experiments were conducted with ROS and Simple Two-Dimensional Robot (STDR) Simulator. Both offline and online scenarios were tested, in offline scenario all obstacles were known, online scenario robot starts with zero knowledge of world and synthetic LIDAR sensor were used to map obstacles.

In simulated scenario with five obstacles Correlation Coefficients and Frechet distances against ground truth were in average 0.98/0.05 for offline SVM and 0.92/0.09 for online SVM

The solution was also implemented succesfully for slightly non-convex obstacles as well

Check out this seminar by Dr. Vela for more details about this work.

Study of Adversarial Attacks on Image Classifiers

Report

Analyzed the effects of popular adversarial attacks such as the Fast Gradient Sign Method (FGSM), Basic Iterative Method (BIM) and DeepFool attack on DNNs trained to classify images on the MNIST and CIFAR-10 datasets.

Trained models on the CIFAR-10 dataset and studied the transferability of adversarial attacks by using examples generated for one model for evaluation by another.

Grad-CAM, a visualisation technique, was used to see where a deep learning model focuses on in an image during evaluation.

For visualisations, models trained on the ImageNet dataset and the FGSM attack were used.

Deep Ensemble Model for Retinal Diseases Detection and Classification

Report

Implemented an ensemble model of DNNs like DenseNets, Resnets, InceptionNets & EfficientNets forretinal fundus images.

Two sets of models were trained - Detector and Classifiers on augmented dataset of total of 3354 images across 28 classes of diseases.

Overall evaluation was performed over 1280 images.

Final ensemble model chieved 0.99 AUROC & mAP.

Protecting Privacy Against Unauthorized Facial Recognition Software

Implemented FAWKES which adversarially cloaks images to protect against facial recognition models.

Used the MTCNN model for face detection.

Implemented and trained the FaceNet and MobileNetV3 models on cloaked images from the VGGFace2 dataset.

MobileNetV3 was able to overcome cloaking with data augmentation, FaceNet model was fooled by FAWKES cloaking

This website is a modification of Jon Barron's website and Priyal Chhatrapati's website. Find the source code to my website here.