Video Highlight


About Me

I am a Principal Applied Scientist at Amazon. I completed my PhD with the Camera Culture group at MIT Media Lab.

I have experience in autonomous vehicles, robotics, computational imaging, computer vision, machine learning, signal processing, optimization, depth sensing, and compressive sensing.

Since completing my PhD, my work has focused on perception, sensing architecture, and end-to-end autonomy for robots and autonomous vehicles.

My PhD at MIT Media Lab focused on computational imaging through occlusions — developing new techniques to see through fog, human tissue, and other occlusions using machine learning, ultrafast cameras, signal processing, and single pixel cameras.

During my undergraduate studies I researched the use of Memristors (a recently discovered fundamental electronic device) for logic in future VLSI circuits and CPU architectures.

Work Experience:

Education:


PhD Projects

Seeing Through Realistic Fog

A technique to see through dense, dynamic, and heterogeneous fog conditions. The technique, based on visible light, uses hardware that is similar to LIDAR to recover the target depth and reflectance.
The system is able, for example, to resolve objects 57 cm away from the camera while the visibility is only 37 cm, demonstrating 6 times better contrast.

Mind Reader

Mind Reader is an online game played against the computer, which is trying to predict the user's next move. The algorithm is based on an online learning framework that learns the user's behavior in real time.
Think you can beat the algorithm? Play the game.

Calibration Invariant Imaging with Deep Learning

Deep learning allows imaging through paper without lengthy calibration.
This method classifies objects hidden behind a scattering layer with a neural network. Training on synthetic data with variations in calibration parameters allows the network to learn a model that doesn't require calibration during real world experiments.

Femto-Pixel – Singe Pixel Lensless Imaging with Ultrafast Compressive Sensing

A time-of-flight based lensless compressive imaging technique that captures scene reflectance with only 50 patterns.
A framework for imaging with single (or few) ultrafast pixels. These are pixels that have pico-second time resolutions and time tag photons as they arrive to sensor.

Imaging Through Scattering with All Photons Imaging

A new method to see deeper and better through tissue with near-infrared wavelengths.
Compared to other methods (e.g. x-ray and ultrasound), using near-visible wavelengths for imaging provides more information about the type of materials they interact with. Our approach is based on time-resolved sensing coupled with an algorithm that is utilizing the scattered photons to overcome the optical scattering.

Robust Movement Imaging in an Endoscopic System

Built on the previous speed imaging technique (see skin perfusion photography below). Here, an endoscopic speed imaging system is demonstrated. Quantitative analysis shows that the suggested global-direct separation technique rejects scattered light which corrupts traditional speed measurement techniques.

Towards In-Vivo Biopsy

A new method to detect and distinguish between different types of fluorescent materials. The suggested technique, with many biomedical applications, provides a dramatically larger depth range compared to previous methods. It uses fluorescent probes, which are commonly used in medical diagnosis since they can be engineered to attach to tissues or change their emission parameters when interacting with specific types of tissues.
This approach may be used to indicate the location of a tumor in deep tissue, and classify it as malignant or benign according to the fluorescence lifetime, thus eliminating the need for X-ray or biopsy. The method is not limited to biomedical applications and can also be used in industrial and general remote sensing applications, such as inspection, identification and flowmetry.

Identi-Wheez: A Device for in-Home Asthma Diagnosis

Demonstrating a child friendly device composed of a stethoscope array. The developed algorithms allow physicians to listen to any point in the lung volume. This will allow parents to diagnose the severity of Asthma at home and alleviate part of the challenges associated with Asthma.

Reinventing the Stethoscope

A device for automatic diagnosis and screening of patients with potential cardiac problems. The device is composed of a synchronized digital stethoscope and an ECG. The two signals allow signal processing algorithms to overcome significant challenges in automatic classification of heart murmurs.
This work started at the MIT Media Lab Engineering Health class and later continued in several REDX workshops in India.

Skin Perfusion Photography

A new computational imaging technique to measure the speed of blood flow in skin tissue (known as perfusion). The system produces speed maps where color indicates the speed of flow. The method is based on a standard camera, a laser source, and a simple algorithm. Its novelty is in a new illumination scheme that enables a robust measurement process.

Single Shot Ultra Fast Imaging

The photography optics of a streak camera is modified to enable single-shot ultrafast imaging. This technique enables capturing videos of non-repetitive phenomena at a trillion frames per second, for example, a plasma discharge from a single laser pulse.

Glasses Free Display

The goal of this project was a display that would correct eye aberration instead of glasses, i.e. a person who usually wears glasses would be able to see information on a display clearly without the need of glasses. The display would show an image that is pre-distorted (for specific eye conditions), but would be perceived as a sharp image by a person standing in front of the display. While other works solved this problem with a combination of hardware and algorithms, here the goal was a pure software solution with a regular display (as a visiting student in Camera Culture group).

LensToy - An Online Optics Design Tool

LensToy is a ray tracing program written in Flash that allows the user to easily manipulate optical elements and design simple optical systems. The original tool was developed by Andrew Adams. Here, additional design capabilities and sensing modalities were added (as a visiting student in Camera Culture group). Download here.

Multi-threaded Switch on Event CPU architecture with Memristors

A new CPU architecture utilizing Memristors as a large and near memory is proposed and analyzed. Having such a large and fast memory enabled in this case a new multi-threading architecture that allows many threads to run efficiently in a long CPU pipeline. The architecture was implemented using Verilog to evaluate its performance (part of undergraduate research project).

Logic with Memristors

Memristors are usually considered memory elements (due to their non-volatile varying resistance).
This project explored ways to implement logic circuits with Memristors. This may allow in-memory computation which is an exciting new computation architecture. Different ways to implement logic using IMPLY gates and MRL gates were explored, as well as a new method called MAGIC (part of undergraduate research project).


Online Talks

Thesis Defense 2019

SPIE Defense + Commercial Sensing 2019

(External link, behind a pay wall)

ICCP 2018

AutoSens 2017

TEDxBeaconStreet 2016


Patents


Publications

For most updated list visit my Google Scholar page.

Journal Papers

Refereed Conference Proceedings (full papers)

Conferences

Thesis

Tutorials


News

Old news, keeping it for historical reference.

Show More


Contact

guysatat (at) alum (dot) mit (dot) edu