I'm a deep learning researcher working on cutting edge deep learning applied to SPECT Imaging in the Radiology Dept. at
UMass Medical School. I hold a Masters degree in Robotics from Worcester Polytechnic Institute where I worked on my thesis under
Prof. Emmanuel Agu for developing a lower extremity wound diagnostics app.
My work Involved studying and mitigating the effects of adverse lighting on
segmentation of wound images using intrinsic image decomposition.
Additionally, I have also been fortunate to work on some more interesting projects
like estimating the egomotion of a vehicle using recurrent convolutional neural networks under
Prof. Riad Hammoud and creating 3D thermal models of indoor environments under
Prof. Shichao Liu
In my parallel universe, I have been exploring the hills of New England, taoist meditation, solving the cube, karaoke singing,
and devouring non-fiction.
Languages: Modern C++, Python, Embedded C
Software/Libraries: Tensorflow, PyTorch, OpenCV, PyData, Weights&Biases, Flask, Slurm, HPC
Hardware: Arduino, Raspberry Pi, Depth and Thermal cameras
Core Competencies: CNNs, LSTMs, Conv-LSTMs, AEs, ML Deployment, Data Science
Akshay Iyer
Worcester, MA 01609
(508) 410-2937
akshay.iyerr@gmail.com
abiyer@wpi.edu
Masters in Robotics • May 2020 •
Related Coursework: Thesis (Deep Learning) | Computer Vision | Artificial Intelligence
Funding/ Scholarships:
• Research Assistantship by WPI
• JN Tata Scholarship by The JN Tata Endowment for the Higher Education of Indians
Publications:
• “Characterizing the effects of adverse lighting on semantic segmentation of wound images and mitigation
using a Deep Retinex Model”, Elsevier Medical Image Analysis, 2020 (In review) (Impact Factor 11.15)
• Reviewer – IEEE Internet of Things Journal (Impact Factor 9.93)
B.Tech in ECE • July 2015 •
Related Coursework: Digital Signal Processing | Probability and Stats | Data Structures and Algorithms
Funding/ Scholarships:
• Sir Dorabji Tata Scholarship for undergraduate studies
Publications:
• "RealNET – Internet of Things in everyday life using Raspberry Pi”, IRJET, 2017 (Impact Factor 7.53)
UMass Medical School • July 2020 - Present
Estimation of a surrogate signal for cardiac motion activity due to respiration:
• Designing a neural network from scratch in the less-ventured area of surrogate signal estimation for motion correction in SPECT imaging
• Performing extensive data analysis to guide experiments to select the optimal network and hyper-parameters
Estimation of attenuation-map from SPECT Emission data:
• Using GAN-based deep learning models to estimate attenuation map from SPECT emission data alone for Boston Children’s Hospital.
• This would eliminate the need for an additional CT scan reducing the exposure to harmful ionizing radiation to children
Worcester Polytechnic Institute • July 2019 - May 2020
Characterizing and mitigating effects of adverse lighting on U-Net segmentation:
• Created the first large-scale dataset of 60k images of a wound moulage under varying lighting conditions and various cameras
• Carried out extensive data analysis and found low light conditions to cause maximum deterioration of UNet performance
• Enhanced the poor lighting separately using intrinsic images and encoder-decoder networks which improved segmentation by 300%
SmartWands Group, WPI • May 2019 - Jul 2019
Image enhancement in adverse lighting:
• Worked on intrinsic decomposition of a single image, an ill-posed problem with high ambiguity and limited training data
• Performed extensive deep learning literature review on intrinsic image decomposition to counteract bad lighting in images
• Trained connected encoder-decoder networks on the ShapeNet dataset to predict reflectance, shape, lighting, and shading
• Used unsupervised reconstruction loss to use unlabeled data for training and generalize to unknown objects and lighting
BOS Lab, WPI • Feb 2019 - Apr 2019
3D thermal modeling of indoor environments:
• Worked under Prof. Shichao Liu to build a 3D thermal map of indoor environments to detect thermal failures in buildings
• Performed depth calibration using Hermann grids and thermal calibration using heated gradient boards
• Performed cross-calibration of a trimodal setup of RGB, depth and thermal cameras
• Developed a new pipeline to create a 3D thermal map using PointCloudFusion
• The pipeline is able to generate 3D thermal maps just from a pair of images
Deloitte Touche Tohmatsu India LLP • July 2015 - June 2018
Robotic Process Automation for UBS, UK
• Automated a highly complex returns process for UBS, UK from end to end, programming in Automation Anywhere
• Worked with business to design the solution document, built code using Automation Anywhere, unit tested the solution and deployed it,
resulting in 83% time savings
Designing India's first Pharma R&D Workbench
• Worked with CIO office of Sun Pharmaceuticals, India's largest pharma company, to design a company wide R&D Workbench
• Worked as the Design-BA for the project as well as was responsible in creating the entire URS document for the workbench
IT Modernization of the Central Bank of Philippines (BSP)
• Created the current state and target state technology architecture to modernize the tech stack at BSP
• Worked on revamping the application architecture and the IT Org Structure for the client
Achieved a fast track promotion at Deloitte (within 2 years vs average of 3-4 years) and rated in the top 5% of employees
Reviewer, IEEE IoT Journal
• On the review board of the IEEE IoT Journal peer reviewing latest paper submissions
in the journal
Workshop Coordinator, Abhilasha Foundation
• Organized a workshop on basics of robotics for underprivileged children at Abhilasha Foundation at
the Nehru Science Center, Mumbai as a part of the CSR event at Deloitte India
Managing Trustee, Smt. Chandra Iyer English School, Mumbai
• Serve as a trustee, overseeing the smooth operations of our family-run school -
Smt. Chandra Iyer English School providing CBSE education at very low fees in the suburbs of Mumbai.
Estimating the egomotion of an autonomous vehicle with live camera feed and a recurrent convolutional neural network. Used FlowNET for optical flow followed by 2 layer LSTM for sequential modeling. Real time performance surpassed Monocular VO KITTI benchmark
Made the Fetch Robot slide a puck to the goal position using Reinforcement Learning. Implemented Vanilla Deep Deterministic Policy Gradient as well as DDPG with Hindsight Experience Replay from scratch
Intrinsic decomposition from a single image is a highly ill-posed problem with high ambiguity and limited training data. Achieved the decomposition to predict reflectance, shape, lighting, and shading from a single image using self supervised learning with an autoencoder
The objective was to make Baxter Robot sort fruits kept on the table in Gazebo. Used HSV segmentation, morphological ops and 3D coordinate estimation in PCL as part of the vision track to identify and classify fruits. Used a Mask RCNN to extend the project to real world fruits.
Involved creation of a 3D thermal model for thermal fault diagnosis of indoor environments. Cross-calibrated a trimodal setup of thermal (FLIR E40)-depth (Orbbec Astra)-RGB (Orbbec Astra) cameras. Developed a new pipeline for the model creation using fusion of point clouds in PCL
Worked on characterizing and mitigating effects of adverse lighting on U-Net segmentation. Created the first large-scale dataset of 60k images of a wound moulage under varying lighting conditions and improved U-Net's segmentation accuracy by 300%
Creating a neural network from scratch to predict surrogate signal for respiratory motion from cardiac SPECT scans.
Implemented sparse-feature based stereo visual odometry that performed in real time using the KITTI dataset. Implemented Lucas-Kanade flow tracking and Perspective-n-Point Algorithm using RANSAC for 3D-2D motion estimation
Performed a comparative analysis of machine learning and deep learning models on the 3-class classification problem of Alzheimer’s detection. Gradient Boosting (67%) and CNN (75%) achieved the highest accuracy, an improvement over similar efforts in the past
Created mixed reality games with MIT Media Lab in the Design Innovation Workshop. Interfaced sensor inputs from physical actions like swiveling and waving to play digital games like Subway Surfers and T-Rex
Created a 5 module Internet of Things based home automation system. Included systems like self watering tweeting plant, smart trash cans, automated courier acceptance, etc. Used Raspberry Pi along with Instapush as the messaging platform.