Page 2
What is Active Sensing?
For some, it’s a hobby!
Page 3
“Object” to be Inferred
Page 6
“Feedback” Guides New “Measurements”
Page 7
“Feedback” Guides New “Measurements”
Page 8
“Feedback” Guides New “Measurements”
Page 9
“Feedback” Guides New “Measurements”
Page 10
“Feedback” Guides New “Measurements”
Page 11
“Feedback” Guides New “Measurements”
Page 12
Background and Motivation
Page 13
Sensors, Sensors Everywhere
Page 14
Integral to Science, Engineering, Discovery
Page 15
Inevitable Data Deluge!
The Economist, February 2010
Page 16
Challenges for Sensing/Processing Systems
Page 17
Active Sensing
-- An Engineering Example --
Page 18
Wide-field Infrared Survey Explorer (WISE)
Need to shield IR (heat) from its own instruments
Sensitive instruments housed in solid hydrogen
Expected lifetime: 10 months!
WISE Mission: (http://www.nasa.gov/mission_pages/WISE/mission/index.html)
“…the infrared surveyor will spend six months
mapping the whole sky. It will then begin a second
scan to uncover even more objects and to look for
any changes in the sky that might have occurred
since the first survey. This second partial sky survey
will end about three months later when the
spacecraft's frozen-hydrogen cryogen runs out…”
Fornax Galaxy Cluster, Feb. 17 2010
Page 19
original signal
(~0.8% non-zero components)
Noisy, non-adaptive sampling
Recovery from adaptive samples
(1/20 “discoveries” are errors)
Recovery from non-adaptive samples
(1/20 “discoveries” are errors)
Noisy adaptive sampling
Astronomical Imaging “On a Budget”
Page 20
Active Sensing for Sparse Recovery
-- Preliminaries and Formalization --
Page 21
DFT
time
Sparsity
DWT
frequency
Page 22
A Sparse Signal Model
signal support set
number of nonzero
signal components
Page 23
Non-adaptive Sampling
Page 24
Alternative: Sequential Experimental Design
Page 25
Alternative: Sequential Experimental Design
Page 26
Alternative: Sequential Experimental Design
Page 28
A Simple Active Sensing Approach
Page 29
Recovery From Non-Adaptive Measurements
Page 30
Fundamentally a Multiple Hypothesis Test
Page 31
Non-Adaptive Support Recovery
Page 32
Adaptive Approach: Sequential Thresholding
Page 35
Probability of Error Bound
(Malloy & Nowak, 2011)
Page 36
greater sensitivity for same precision budget or lower
experimental requirements for equivalent sensitivity
Improvements Through Sequential Design
Page 37
Active Sensing for Sparse Recovery
-- A Relaxed Error Criteria --
Page 38
Measuring Error: False Discoveries
Here, FDP = 3/5
Page 39
Measuring Error: Non-Discoveries
Here, NDP = 5/7
Page 40
Limits of Non-adaptive Sampling
Page 41
Signal Strength
Sparsity
Estimation Possible
(FDP + NDP 0)
Estimation Impossible
(FDP + NDP 0)
Sharp Delineation in “Parameter Space”
Page 42
What to Do in Low SNR Settings?
Page 43
Idealized Example
Page 44
Distilled Sensing (DS)
Page 45
DS
(k=4)
Estimation Possible(FDP + NDP 0)
Estimation Impossible(FDP + NDP 0)
Non-adaptive limit
Estimation Possible(FDP + NDP 0)
Achievable using DS
Equal Allocation of Sensing Resources
Page 46
Estimation Possible(FDP + NDP 0)
Estimation Impossible(FDP + NDP 0)
DS
(k=6)
Estimation Possible(FDP + NDP 0)
Non-adaptive limit Achievable using DS
Equal Allocation of Sensing Resources
Page 47
original signal
(~0.8% non-zero components)
Non-adaptive observations
Adaptive recovery
(FDP = 0.05)
Non-adaptive recovery
(FDP = 0.05)
DS observations (k=6)
Recall: Noisy Astronomical Imaging
Page 48
(FDP + NDP 0)
(FDP + NDP 0)
(FDP + NDP 0)
Non-adaptive limit DS, k=8
Unequal Allocation of Sensing Resources…
Page 50
The Curse of Dimensionality…
Page 51
…and the Virtue of Adaptivity!
Page 52
Active Sensing for Sparse Recovery
-- Adaptive Compressive Sampling --
Page 53
Improvements w.r.t. Other Resources?
Page 55
Curse of Dimensionality: Compressed Sensing
Page 56
…and the Virtue of Adaptivity!
Page 57
Generalization Beyond Gaussian Models
-- Sequential Thresholding --
Page 66
End of Presentation…