Associate Professor, MIT Media Lab
Ramesh Raskar is founder of the Camera Culture research group at the Massachusetts Institute of Technology (MIT) Media Lab and associate professor of Media Arts and Sciences at MIT. Raskar is the co-inventor of radical imaging solutions including femto-photography, an ultra-fast imaging camera that can see around corners, low-cost eye-care solutions for the developing world and a camera that allows users to read pages of a book without opening the cover. He is a pioneer in the fields of imaging, computer vision and machine learning.
Raskar’s focus is on building interfaces between social systems and cyber-physical systems. These interfaces span research in physical (e.g., sensors, health-tech), digital (e.g., tools to enable keeping data private in distributed machine learning applications) and global (e.g., geomaps, autonomous mobility) domains. Recent inventions by Raskar’s team include transient imaging to look around a corner, a next-generation CAT-scan machine, imperceptible markers for motion capture, long-distance barcodes, touch + hover 3D interaction displays and new theoretical models to augment light fields to represent wave phenomena.
Raskar has dedicated his career to linking the best of the academic and entrepreneurial worlds with young engineers, igniting a passion for impact inventing. Raskar seeks to catalyze change on a massive scale by launching platforms that empower inventors to create solutions to improve lives globally.
Raskar has received the Lemelson Award, ACM SIGGRAPH Achievement Award, DARPA Young Faculty Award, Alfred P. Sloan Research Fellowship, TR100 Award from MIT Technology Review and Global Indus Technovator Award. He has worked on special research projects at Google [X] and Facebook and co-founded and advised several companies. He holds more than 80 US patents.
Making the Invisible Visible: Within Our Bodies, the World Around Us, and Beyond
23. Optical Jumbled Brush Endoscope
Heshmat, Nature SciRep16
Cellular resolution at 5mm
NSF Moonshot
FLIM Location behind Tissue
Satat, Nature Comm 15
Satat, Nature SciRep 17
CT-scan in a Rickshaw
Kadambi 17
24. 1mm
Ballistic Limit
Surface
100 um
1 mm
10 mm
100 mm
10 um
Spatial Resolution
1 um 10 um 100 um
OCT2/3 photon
Confocal
Computational
Photoscatterography
Depth
‘All Photon’ Imaging for Tissues
NSF Expedition
2017-2022
26. Beat Diffraction
Gated imaging to overcome ambient light
‘Negative light' via destructive interference inside any volume
Focus at or ‘heat’ any voxel ..
Conquer time ..
• Seeing around corners
• Fog/Closed book
• Endoscopes/ Optical Brush
• Fluorescence Lifetime
28. Tancik, Satat, Raskar, Flash Photography for Data-Driven Hidden Scene Recovery28
Training data = Renderings
100k’s renderings, no real photos
Transfer to real experiments
38. 38
Spot Diagram
on LCD
Inverse of Shack-Hartmann wavefront sensor
User interactively creates the Spot Diagram
1. Displace 25
spots with
smart UI
CellPhone
LCD
EyePiece
2. Displace spots
till single dot
perceived
39. EyeSelfie: Retinal Self-Imaging, Swedish et al, SIGGRAPH 2015
Raskar TEDMED 2013
Roesch, Swedish, Raskar, Clinical Ophthalmology 2017
44. Pool ‘small’ data
+ Train ML
Private ML
No Exchange of
Raw Patient Data
Gupta, Raskar ‘Distributed training of deep neural network over several agents’, 2017
Server
45. a. Automating ML
Published in International Conference on Learning Representations (2017)
Baker, Gupta, Naik, Raskar, ICLR 2017
Teacher: RL Student: Supervised ML
48. Split Learning (MIT)
~Federated Learning
Share Wisdom.
Not Raw Data
Data
Utility
Train Model
EncryptSmashObfuscate
Add Noise e.g.
Differential Privacy
Homomorphic
Encryption
Data
Protect
Infer
Statistics
Anonymize
Training Deep Networks without exchanging raw patient data
52. 52
VGG over CIFAR 10 ResNet over CIFAR 100
Federated
Split
Compute Bandwidth
Vepakomma, Swedish, Gupta, Dubey, Raskar 2018Raskar, McMahan et al CVPR 2019
53. Absolute
Value
Relative
Value
Conditional Value,
additional users or features
Intrinsic
Goal-independent Value independent of the final goal
Goal - specific Value based on ML algorithm
Privacy-preserving Value without revealing raw data (with or without goal)
Extrinsic
Supply-Demand Speculated value using game theoretic multi-party interests
Vepakomma, Swedish, Raskar 2019
c. Data Markets for ML