Sensing Your Health

1
Sensing Your Health

• Roozbeh Jafari
• Electrical Engineering Department
• UT Dallas

2
Outline

• Medical Embedded Systems
• System Architecture for Body Sensor Networks
• Pilot Application
• Physical Activity Monitoring
• Sport Training Systems
• Conclusion

3
Why Should You Care?
4
Medical Embedded Systems
5
Embedded systems

• Computing system are everywhere
• Embedded within electronic devices
• General purpose vs. embedded computing
• Very broad to define! Nearly any computing system
  other than a desktop computer
• Designed for a particular (or class of)
  applications
• Constrained in terms of computing, communication,
  etc.
• Billions of units produced yearly, versus
  millions of desktop units

6
Medical Embedded Systems

• For medical monitoring and diagnostic
• Direct concerns
• Cost-effective
• Mobile
• Lightweight/ low profile/ comfortable
• Power aware
• Indirect concerns
• Ease of use
• Adaptable/ customizable to meet individuals need
• Reliable, robust and secure
• Work with various external and environmental
  sensors

7
BSN System Architecture
8
Body Sensor Networks
Applications fall monitoring for elderly,
rehabilitation, sports medicine, gait analysis,
and locomotion monitoring, posture detection
Sensors inertial sensors, skin conductance
sensors, EMG, ECG, piezoelectric, flex, pressure
sensors

• Action recognition, quantitative and qualitative
  measures on human actions (i.e. balance,
  consistency, coordination)
• Normative study
• Database (Google for human actions) Data
  mining, Indexing, Scoring
• Modeling Compact models for human actions,
  Grammar construction
• Design and optimization techniques

• Applications sport training systems and virtual
  coach golf, baseball, stress evaluation for
  soldiers

Body Sensor Networks
9
System Components

• Sensors
• 3-Axis accelerometers
• Gyroscopes
• Flex
• Pressure
• Piezoelectric film (floor sensors)
• Motion capture
• EMG
• Galvanic skin response
• Electronic textile sensors

Courtesy of Bruce Gnade
10
Sensors on Body
11
System Components

• Processing units Telos motes
• Devices that incorporate communications and
  processing into a small package
• TI MSP430 microcontroller
• Analog/digital interface
• USB and serial connection
• IEEE 802.15.4 radio
• Nonvolatile storage (1MB)

Courtesy of moteiv
12
Constraints

• Wireless communication
• Multiple motes send to a single central node
• Over an unreliable wireless channel
• Limited processing power
• 8MHz, 10k RAM
• Power constraints
• 2 AA Batteries

Courtesy of moteiv
13
Signal Processing
14
Platform Design
15
Objective

• Enhance wearability and convenience while
  guaranteeing signal processing requirements
• Large variation in terms of patient preferences,
  environmental factors, nature of physical daily
  activitiesetc

16
Resources

• Sensors
• Processing units
• Battery
• Transceivers
• Interconnections
• Wires
• Wireless medium

17
Wearability Factors

• Physical Factors
• Size
• Weight
• Heat dissipation
• Light emission
• Sensor location (on body)
• Environmental Factors
• Environment
• Daily activities

18
System Wearability
19
Example
Actuation
Sensing
Processing
Battery Powered Sensor Node
Gateway
(1)
Sense Transmit
Receive Process Actuate
Gateway
Battery Powered Sensor Node
(2)
Sense Process Transmit
Receive Actuate
20
Physical Movement Monitoring
21
Objective

• Monitoring physical movements using distributed
  motion sensors
• Why important?
• reducing the reliance on institutionalized health
  care and moving towards the adoption of a home
  telecare paradigm for aging population

22
Application Scenarios

• Fall monitoring
• Rehabilitation
• Sports medicine
• Gait analysis
• Locomotion monitoring
• Posture detection
• Sports training systems

23
Experimental Setup for Daily Physical Movements
(eight sensors)

• 8 sensors
• 10 trials over 25 movements

24
List of Movements (eight sensors)

• 1. Stand to sit
• 2. Sit to stand
• 3. Stand to sit to stand
• 4. Sit to lie
• 5. Lie to sit
• 6. Sit to lie to sit
• 7. Bend
• 8. Kneeling, right leg first
• 9. Kneeling, left leg first
• 10. Turn clockwise 90 degrees and turn back to
  the initial position
• 11. Turn counter clockwise 90 degrees and turn
  back to the initial position
• 12. Turn clockwise 360 degrees
• 13. Turn counter clockwise 360 degrees
• 14. Look back
• 15. Move forward (1 step)
• 16. Move backward (1 step)
• 17. Move to the left (1 step)
• 18. Move to the right (1 step)
• 19. Reach up (to a cabinet) with one hand

25
Raw Sensor Readings (eight sensors)
One Action
X Accel
Y Accel
Z Accel
X Gyro
Y Gyro
Stand to sit 10 trials for one subject
26
Raw Sensor Readings (eight sensors)
One Action
X Accel
Y Accel
Z Accel
X Gyro
Y Gyro
Look back over right shoulder 11 trials for one
subject
27
Raw Sensor Readings (eight sensors)
One Action
X Accel
Y Accel
Z Accel
X Gyro
Y Gyro
Reach up with two hands 10 trials for one
subject
28
Experimental Setup (three sensors)
29
Experimental Setup (three sensors)
30
Sequence of Movements (three sensors)

• Initial position lying
• 1. sit and stand
• 2. walk to a desk
• 3. sit at the desk (swivel chair)
• 4. start writing
• 5. stand
• 6. walk to a chair (armchair1)
• 7. sit, relax
• 8. stand
• 9. Walk to the shelves
• 10. grasp a plate (two hands)
• 11. walk to a dinning table
• 12. bend and leave the plate on the table
• 13. sit and start eating (dinning chair)
• 14. stand
• 15. Walk to a coffee table
• 16. sit (armchair2)
• 17. pick an object from the table and leave it on
  the table

31
Raw Sensor Readings (three sensors)
X Accel
Y Accel
Z Accel
Complete sequence
32
Raw Sensor Readings (three sensors)
One Action
X Accel
Y Accel
Z Accel
Writing
33
Locomotion Monitoring
34
Objectives

• Quantify coordination and consistency of walking
• Study the development of locomotion in infants
• Diagnose disorders such as autism

35
Measures

• Coordination
• Compare frequency components and phase offsets of
  different joints
• Consistency
• Compare the frequency response of two points in a
  signal as a function of the distance between
  those two points

36
Coordination

• Difference between main frequency components of
  two joints

Adult Walking
Child Walking
Main frequency components differ slightly
Main frequency components match
37
Consistency

• A repetitive movement like walking is consistent
  if each cycle of the movement is similar
• Consistency function
• S is the PSD of the signal
• Measures the similarity of the signal to the same
  signal at a different point
• Exhibits periodicity when the original signal is
  consistent

38
Consistency
Walking on balance beam inconsistent
Walking on floor consistent

• Time between adjacent minima is period of
  original signal
• Value at minima corresponds to how similar
  different cycles are

39
Experimental Setup

• Seven movements, all variations of walking
• 26 subjects (so far)
• Ages 19-22
• Twelve sensors

40
Experimental Setup
41
Walking Variations

• Designed to present different levels of
  difficulty so subjects exhibit a range of
  coordination and consistency
• 1. Normal 5. Slowly
• 2. Backward 6. 100 steps in a circle
• 3. Without shoes 7. Balance Beam
• 4. Quickly

42
Analysis

• Objective
• Determine baseline for coordination and
  consistency for normal adult walking
• Consistency and coordination values calculated
  for all trials
• Mean and median calculated for each movement

43
Consistency Results
1. Normal 2. Backward 3. Without shoes 4.
Quickly 5. Slowly 6. 100 steps in a circle 7.
Balance beam 8. Balance beam (steps normalized
in time)
44
Raw Signals Walking Normally (subject 24)
45
Coordination Between Joints Walking Normally
(subject 24)
46
Raw Signals Walking in a circle (subject 24)
47
Coordination Between Joints Walking in a circle
(subject 24)
48
Coordination Results

• Right arm vs. left ankle

1. Normal 2. Backward 3. Without shoes 4.
Quickly 5. Slowly 6. 100 steps in a circle 7.
Balance beam
49
Results

• Consistency

• Coordination

50
Comparison to Expectations
51
Future Work

• Refine measurements to better match expectations
• Gather data from more subjects
• Analyze effect of subject demographic on
  consistency and coordination

52
Baseball Swing Trainer
53
Motivation

• Private coaches are expensive
• Cost increases based on time
• Alternative virtual coaching system
• Automated, accurate training system that gives
  fine-grained coaching feedback
• Monitor movements
• Analyze movements correctness
• Provide information on how to improve

54
Sensor Placement

• Sensor placement (right-handed batters)
• Locations bat barrel, bat grip, right wrist,
  left shoulder, waist, and both ankles
• Chosen to identify potential swing problems

55
Body Choreography Modeling

• Swing analysis is based on body choreography
  modeling
• Actions, such as a swing, consist of a sequence
  of motion primitives
• Each body part has a sequence of primitives
• Timing between body parts is important

56
Body Choreography Modeling

• Problems with an action occur if
• Sequence is incorrect
• Timing within a sequence is wrong
• The swing should be three times as long as the
  windup
• Sequences for different body parts are misaligned
• Shoulders should rotate at the same time as the
  hips

57
Motion Primitives

• Motion primitives are atomic bits of motion
  analogous to phonemes in speech
• Potential motion primitives
• Moving the foot forward
• Swinging in a slight arc
• Fast forward swing
• Slow forward swing

58
What Is a Good Swing?

• Our definition based on the Ted Williams swing
• Begin with neutral stance
• Cock hips (windup)
• Step into the pitch, rotating the front foot
  toward the pitcher
• Rotate the hips and swing the bat
• Follow through without rolling wrists

59
Preliminary Experiments

• We looked at the following problems
• No hip rotation
• No front foot step
• No rotation during front foot step
• No back foot rotation during swing
• No windup
• Wrist roll

60
Preliminary Experiments

• Seven movements
• One perfect swing
• Six swings with only one of the problems in the
  previous slides present
• Subjects 2 males, 1 female, ages 23-30
• 20 trials (swings) per swing type

61
Building Motion Transcripts

• Motion primitives found using k-means clustering
  trained on the first 10 trials
• Used k7
• A set of clusters was independently chosen for
  each sensor
• Cluster centroids define primitives
• Samples from all trials assigned to closest
  centroid

62
Features for Clustering

• Features extracted from a 5-point window centered
  on the sample
• Features include
• Standard deviation
• Mean
• Max mean
• Min mean

63
Example Transcript
64
Explanation of Transcript

• Motion primitives represented by different colors
• Parts of swing identified from video
• The sensor representing each action was chosen by
  deduction and trial and error

65
Swing Components

• Above are sequences identifying different swing
  components
• TP positive recognition in the Perfect swing
• FP positive recognition in the swing that did
  not include this
• We did not examine in-sequence timing and
  between-sequence alignment
• That would reduce high FP

66
Further Examples
67
Future Work

• Pick better features for clustering
• Investigate structured method to select number of
  clusters
• Research other clustering techniques
• Hierarchical
• Mixture of Gaussians
• Analyze timing between primitives on different
  nodes

68
Golf Swing Trainer
69
Objective

• Design a real-time wearable Golf Training system
  that provide qualitative feedback

70
Golf Swing Model

• Golf swing as a four-phase motion

Follow Through
Down Swing
Take Away
Back Swing
71
Definition of Each Segment

• Tack-Away
• starts after address position and ends when club
  is parallel to address line
• Back-Swing
• Follows take-away, continues until club is lifted
  to its highest point behind the player.
• Down-Swing
• Follows back-swing, the club is brought back down
  to hit the ball.
• Follow-Through
• After impact, brings the club to highest point in
  front of the player.

72
Significant Parameters

• In-plane swing
• Proper swing lies in the plane created by the
  golf club and the ball.
• Circular motion
• Hands are swung in a circle about a point in the
  upper part of the chest.

73
Swing Plane
74
System Setup

• 5 sensor nodes
• 2 on golf club, 1 on forearm, 1 on left arm, 1 on
  back
• 2 subjects, ages between 25 and 35

75
Sample Waveforms
Good Swing
76
Sample Waveforms
Out of plane Swing
77
Major Segments of a Swing

• Accelerometer readings, node placed on the golf
  club

78
Experiments for In-plane Swing

• Motions
• One in-plane (good) swing vs. four different
  variations of out-of-plane (bad) swing
• 2 subjects, 10 trials each motion
• k-NN Classifier
• 50 training 50 test
• Features
• Peak-to-peak amp
• Start-to-end amp
• RMS power

79
Classification Results (in-plane)

• Accuracy
• Good vs. bad back-swing, 90
• Good vs. bad down-swing, 100
• Good vs. bad follow-through, 94

80
Experiments for Wrist-rotation

• Motions
• One In-plane (good) swing vs. six different
  variations of wrist rotation
• Bad swings 3 clockwise 3 counter clockwise
• 2 subjects, 10 trials each motion
• k-NN Classifier
• 50 training 50 test
• Features
• Peak-to-peak amp
• Start-to-end amp
• RMS power

81
Classification (wrist rotation)

• Accuracy
• Good vs. bad take-away, 95.7
• Good vs. bad back-swing, 95.7
• Good vs. bad down-swing, 97.1
• Good vs. bad follow-through, 91.4

82
Qualitative feedback

• Objective
• How bad a bad swing is?
• Suspicion
• Qualitative metric is proportional to the quality
  of the motion

83
Observations

• LDA projection of training set into 2D space for
  wrist rotation motions

84
Observations

• Projection of bad swings for a subject performing
  two wrist-rotated swings

85
Methods

• Feature selection
• Find features contributing to discrimination
  among different variations of a bad swing
• Discriminant functions
• Find linear projection of feature space (e.g.
  Linear Discriminant Analysis LDA) on training
  data
• Qualitative feedback
• Project each test pattern using discriminant
  function

86
Techniques for wrist rotation

• Stepwise disctiminant analysis to find
  significant features
• Each segment has its own feature set
• Technique applied to clockwise and counter
  clockwise rotations
• Extract first discriminant function for each
  segment/rotation
• First projection gives well-separated classes
• Validate the technique

87
Experiments (wrist-rotation)

• Training set
• 2 subjects, 3 variations of clockwise, 3
  variations of counter clockwise, 10 trials each
• Test set
• 2 subjects, clockwise and counter clockwise, 2
  variations each
• System was tested on patterns of individual
  subjects
• Initial feature space
• 20 different statistical and morphological
  features collected from 5 nodes, each 5 sensors

88
Sample projections

• Down-swing, Counter clockwise, subject 1

89
Sample projections

• Down-Swing, Counter clockwise, subject 2

90
Sample projections

• Back-swing, Clockwise, subject 1

91
Sample projections

• Back-swing, Clockwise, subject 2

92
Technical Questions?

• What sensors?
• Where to placed them?
• When need to be active?
• Modeling aspects of signal processing?
• Customizability
• One may choose to wear the sensor on wrist while
  the other as a necklace
• Inconsistency in dimensionality of data?
• Calibration

93
Action Coverage
94
Class Distribution
95
Compatibility Graph
96
Sample Compatibility Graph(real data, 12
movements, waist node)
3
6
Turn Clockwise 90 degrees
7
8
10
14
15
17
Look back clockwise
20
22
23
Grasp, Turn (90), Release
25
10, 14 and 22 are not compatible!
97
Example (Action Coverage)
2 nodes instead of 4
98
Problem Complexity

• Set Cover problem
• NP-hard
• Best approximation factor O(log n)

99
Class Separability

• Using Bhattacharyya distance
• Separated classes have distance 2

100
Greedy Approximation
At each step, pick the graph that has the largest
number of uncovered edges
101
ILP Formulation
102
Dynamic Optimization

• Pick a node as Master, e.g. node I

• Find missing edges incident with target class ?
  edge (A,B)

• Pick complementary node/s

103
Experimental Setup

• 8 sensors
• 3 subjects, average age 29
• 10 trials over 25 movements

104
Categorized Movements
105
Classification Accuracy

• 30 used for training
• 70 for validating proposed techniques
• K-NN classifier

106
Results for 25 movements

• 5 nodes reported by ILP solver (2,4,5,7,8)
• 6 nodes by greedy algorithm (2,3,4,5,7,8)

107
Classification Accuracy
108
Compatibility Graph (Applications)

• Action Coverage
• Minimize number of active nodes while maintaining
  high classification accuracy
• System Lifetime
• Maximize system lifetime by picking different
  subsets providing full coverage

109
System Lifetime
110
System Lifetime

• Maximize lifetime by picking a subset of sensor
  nodes that provide full coverage.
• Assumption Subsets providing full coverage are
  given.

111
LP Formulation
112
Results (node activation)
113
Results (performance)
114
Results (scheduling)

• 8 nodes 4 movements 3 subjects
• Lifetime 204 hours 3 times a single subset
  selection

115
Grammar Construction
116
Motivation

• Inherent characteristics
• Movement and spoken language use similar
  cognitive substrate in terms of grammatical
  hierarchy
• Distributed recognition
• Ability of each node to recognize movements
  varies based on the type of movement and physical
  placement of the node

117
Objective

• Design a Linguistic model for human movement in
  BSNs
• Advantages
• Compact representation of movements
• Extract temporal characteristics of motion
  (complex movements human behaviour)

118
Design Methodology

• Phonology
• Find basic primitives (phonemes) and assign
  appropriate symbols to them
• Morphology
• Combining primitives to form a higher-level
  movement (word)
• Syntax
• Constructing behaviour (sentence) from predefined
  rules

119
Primitive Construction Detection
120
Methods

• Group similar movements using clustering
  techniques (e.g. k-means)
• Remove low-quality clusters
• Each cluster corresponds to a primitive
• Label each movement in terms of its primitives
• Construct a decision tree for action recognition
• Classify a test pattern using decision tree

121
Phonology (example)

• 3 nodes 4 movements 7 primitives

Movement A
Node 3
Expressing movements in terms of primitives A
P1, P4, P6 B P2, P4, P6 C P2, P5, P6 D
P3, P5, P4
122
Phoneme Identification (example)
s1 (A,B) s2(C) s3(D)
s1 (A,D) s2(B,C)

123
Decision Tree

• How to construct optimal tree?

?
Optimal Decision Tree
124
Static Decision Tree

• Optimal Decision Tree is equivalent to a linear
  ordering of sensor nodes s.t. cost of recognition
  is minimized
• Equivalently, number of sensor nodes in linear
  ordering must be minimized

125
Min. Cost Identification

• MCI is the problem of finding a complete linear
  ordering of sensor nodes such that cost of
  ordering (time to converge) is minimized
• MCI is NP-Complete (by reduction from Min. Sum
  Set Cover)
• Best approximation factor is 4

126
Definition (LDS)

• Ability of each node is encoded in LDS, a set of
  movement pairs discriminated by the node.

• Local Discrimination Set
• LDS1(A,B), (A,C), (A,D), (B,D), (C,D)
• LDS2(A,C), (A,D), (B,C), (B,D)
• LDS3(A,D), (B,D), (C,D)

127
Greedy Approximation
128
Experimental Setup

• 18 nodes
• 30 movements
• 10 trials each
• 3 subjects
• 1 male, age 32
• 2 females, ages 22 and 55

129
Movements
130
Ordering

• 10 sensor nodes reported by greedy algorithm
  (maximum of required nodes)

131
Results (graph view of ordering)
132
Classification Accuracy

• Features
• Peak-to-peak amp
• Start-to-end amp
• RMS power
• Mean Value
• Standard Deviation
• Constructed static decision tree
• Identification accuracy using 10 nodes structure
  91

133
Ongoing research

• Constructing full decision tree for dynamic
  optimization (faster convergence)
• Combining primitives to find higher level
  movements (human behavior)

134
More Technical Problems
135
Distributed Signal Processing
136
Challenge

• Automatic segmentation and annotation
• Large and distributed feature space!
• Feature conditioning
• Distributed classification
• Expensive communication

137
Segmentation - Annotation
138
What is Point Annotation?

• Point annotation locates and labels specific
  points of interest
• Points are identified by considering a small
  region surrounding the points
• Amplitude, shape, and context may all be helpful
• Segments may be formed from the interval between
  two specific points

139
Examples of Point Annotation

• GAIT - Walking
• Walking is a continuous motion that can be
  cyclically divided into three segments
• EKG
• An EKG contains P, Q, R, S, and T waves
• The relative timing and amplitude are significant
• Sports
• The swings of a club, bat, or racket can be
  divided into phases

140
Classification Approach

• Extract features from a moving window
• Use a classifier to determine if
• An event has occurred
• The label of the event
• Events tend to be sparse, so there are far more
  non-events than events

141
Pilot Application

• 8 sensors on each subject
• Sensors have 3 axis accelerometer and 2 axis
  gyroscope
• Placed on all major limbs
• 3 subjects performed 25 transitional movements
  with 10 trials each
• Movements included
• Sit-to-stand, sit-to-lie, turn counter clockwise
  90 degrees, jumping

142
Goals

• We want to identify
• Start of the action.
• End of the action
• Neither start nor end
• We exclude a region around start and end points
  equal to the window size

Stand to Sit
143
Preliminary Results

• We present the results using linear discriminant
  analysis
• Red is non-events, green is start, blue is end

All Movements, 10 pt window
All Movements, 16 pt window
All Movements, 24 pt window
144
For Individual Movements

• What if we know the movement?
• Once we identify the action, we may want to
  identify specific parts of it
• This eliminates the confusion with similar
  signals from other movements
• Toy applications
• Identifying elements of a swing in sports
• Partitioning GAIT into specific segments

145
Preliminary Results

• Events are much better separated
• All use a 10 point window

Kneeling, right leg first
Move forward one step
Sit to Stand
146
Morphological Features

• The signals around the points of interest tend to
  have characteristic shapes
• We can use features that recognize the shapes
• A simple method
• Use every point in a window around the event as a
  feature
• Can we use a smaller number of features?

147
Minimal Basis Functions

• The shape for each event i can be represented by
  a function
• Example functions
• This shape can be approximated by using k
  orthonormal basis functions

148
Properties of Orthonormal functions

• The two properties of orthonormal function

149
Basis Function

• Perfect representation
• Achieved using each shape in the training set as
  a basis
• Or by using a function for each sample in the
  window
• Perfect representation is undesirable
• Too many basis functions/features
• An approximate representation may actually help
  remove noise from the shape

150
Minimal Least Squares

• Least squared error for all events is
• For given basis functions, we minimize error by
  (take derivative of above function and set to 0)
  

151
Finding the Basis Functions

• With zero mean this equation exactly matches PCA
  for a matrix where
• Rows are events
• Columns are samples in the window around the
  event
• The first k eigenvectors from PCA are the basis
  functions that minimize error

152
Results

• Here we show
• Original signals
• Eigenvectors from PCA
• Reconstructions using the first k eigenvectors,
  with several k

Even with only 2 basis functions, the
reconstruction is very good
153
Results for multiple classes
Five eigenvectors do a very good job of
recognizing shape, and represent a 75 decrease
in necessary features from 24 for the whole
window.
154
Continuing Research

• Annotate while minimizing inter-node
  communication
• Use LDA to generate basis functions that are best
  a discrimination, not representation
• Explore graph-based methods of fusing mote
  knowledge

155
Constant Communication
156
Problem

• Wireless communication is expensive on sensor
  nodes
• Limited power
• Limited bandwidth

156
157
Objective

• Develop a method in which the communication cost
  at each node is fixed and small
• Each node sends k units of information
• Choose these k units such that classification
  accuracy is maximized

157
158
Choosing Information to Send

• Requires global knowledge of the system
• Some movements require information from more than
  one node
• Information at one node may be redundant with
  information at another
• But each node must make a local decision about
  what to send

158
159
Choosing Information to Send

• Data has high dimensionality
• For eight sensors, 240 features
• Use LDA to determine the significance of features
• LDA projects the data onto vectors so that the
  distance between classes is maximized
• Each vector given by LDA is some combination of
  the original features

159
160
Results of LDA
Top seven features for first dimension given by
LDA
Projection of 25 classes into first three
dimensions from LDA
25 classes in first three dimensions of original
feature space
160
161
Communication Model
162
Objective

• Build a model of communication based on notion of
  compatibility graph
• Encode local knowledge within the model
• Include cost of information per node

163
Intuition

• Encode information within compatibility graphs in
  a flow network
• Each movement corresponds to a path in the
  network
• Each node is a link in the flow network
• Costs are associated with sensor nodes.

164
Example

• 3 mvts (A,B,C) 3 nodes (n1,n2,n3)
• Links corresponding to senor nodes are shared
  among different paths (motions) (blue region),
  but each movement has its own pink area.

A sub-graph showing information regarding
movement A. System requires 2 units of
information to detect A. Demand would be 2
units of flow for commodity A (As, At)
165
Problem

• Mutli-commodity Min-Cost flow problem
• Paths correspond to movements
• Demands are determined based on units of
  information required to detect each movement
• Flows will show how information must be sent.

166
Class Resource Similarity (CRS)
167
What characterizes CRS?

• If two classes are similar
• They both can be classified using a nearly
  identical feature set with similar priorities.
• Therefore the same resources are used for each

168
Group Similarity

• This can be extended to a group of classes
• The minimum pair-wise similarity between classes
  in the group meets a minimum threshold
• A graph can be formed where a connection between
  two classes indicates that the threshold has been
  met
• A clique represents a group by the above
  definition

169
Calculating Similarity

• We need to know the relevance of each feature to
  a class
• Relevance is how good a feature is at uniquely
  distinguishing a class
• Mutual Information is frequently used in the
  literature

170
Similarity

• For each class, construct a feature vector
  containing each feature and its relevance to the
  class
• Use cosine of angle between feature vectors for
  two classes

171
Redundancy

• If a feature is duplicated 6 times, then it will
  appear to be six times as important.
• Most feature selection schemes include ways of
  eliminating or reducing redundant features
• However, if we eliminate redundant features, two
  classes whose second-best feature sets are
  identical, may appear totally dissimilar due to
  their first choices.

172
Redundancy

• So we must intelligently reduce the relevance of
  redundant features
• Omega is the reduction factor
• Rho is the pearson correlation coefficient
• Square root ensures unbiased voting for
  correlated features

173
Grouping Criteria

• Overlap up to a point is good
• Large groups are good
• Even coverage is nice
• Complete coverage is required

174
Grouping First Try

• Listed all cliques such that no clique is a
  subgraph of another clique.
• Developed an algorithm
• For 25 classes, and an aggressive cutoff, there
  were 9864 such cliques, with the max-clique
  containing 14 classes

175
Grouping Second Try

• Filtered the first results using an algorithm
  designed to meet the criteria.
• Results 8 groups. The first group has 14
  members, the remaining 7 each have two (there is
  some overlap)
• This is mostly because there arent any other
  cliques that are dissimilar to the first that
  have more than two members

176
Acknowledgment

• Hassan Ghasemzadeh
• Eric Guenterberg
• Jaime Barnes
• Katherine Gilani
• Danny Tseng
• Cale Dingman
• Hanan Abdulahi

177
Acknowledgment

• Nisha Jain
• Chellappan Valliyappan
• Jake Wurzer

178

• For more information
• www.essp.utdallas.edu