TRADING%20OFF%20PREDICTION%20ACCURACY%20AND%20POWER%20CONSUMPTION%20FOR%20CONTEXT-AWARE%20WEARABLE%20COMPUTING

1
TRADING OFF PREDICTION ACCURACY AND POWER
CONSUMPTION FOR CONTEXT-AWARE WEARABLE COMPUTING

• Presented By Jeff Khoshgozaran

2
Motivation

• Wearable devices sensing user information
• Context-Aware Mobile Computing
• Previous work
• Power consumption in full power mode
• Quickly depletes a critically constrained
  resource
• High sampling rate to provide accuracy
• Computational and space-intensive solutions
• Lack of scalability for knee and hip-worn sensors

3
eWATCH

• A context-aware wearable platform
• Several sensors including two-axis accelerometer
• Three power states for sensing and classifying
  data
• Full Power
• Active CPU, active peripherals ( 30 ms)
• Idle State
• Core clock turned off, active peripherals
• Waiting for the next sample
• For SR6 Hz, time interval166ms
• Low Power
• State is active most of the time
• Inactive CPU and peripherals
• Active real-time clock
• Scheduling next wake up
• Using selective sample algorithm

www.chronicle.pitt.edu, http//www.cmu.edu/
4
Time/Frequency-Domain-Based Classification

• 5 second windows for computing features
• Time-based
• Using means, variances, median, etc.
• Frequency-based
• Using FFT on values of both accelerometer axes
  separately

Human movement is periodic Frequency-based
Approaches good for classifying accelerometer
data Less expressive for very low sampling rates
http//www.seas.gwu.edu/ayoussef/cs225/
5
Battery Lifetime Vs. Sampling Rate
While important, computation is not the dominant
factor in reducing energy consumption
More costly computation and more dimensions
6
Using SVM for Classification

• A multi-class SVM used for actual classification
• Detect and exploit complex patterns in data
• Good for representing complex patterns
• Good for excluding unstable patterns (
  overfitting)
• Computationally expensive training
• Very efficient classification (hardware friendly)
• Guassian Radial Basis Function used as kernel to
  classify non-linear data
• The class of kernel methods implicitly defines
  the class of possible patterns by introducing a
  notion of similarity between data
• Implicit and non-linear embedding of data in
  high-dimensional spaces

Separated by a hyperplane in feature space
7
Power-optimized Classification Experiments

• Training data captured by 3 test participants
• Each activity recorded for 10 minutes
• Data was split into different recorded activities
• Data was partitioned into blocks of 5 seconds
• Used to extract time/frequency domain features
• Labeled examples used for training multi class
  SVM
• Prediction accuracy power consumption computed

8
Results
85 Increase
Optimum sampling frequency of 6 Hz
For all but extremely low frequency ranges,
frequency based features perform superiorly.
9
Selective Sampling vs. Prediction Accuracy

• What Further reduce energy consumption
• How Selective Sampling
• Why Human activity a continuous process
• Person more likely to continue an activity than
  to change to another at a point in time
• Selective sampling schedules classification
• Reduces number of observations
• Saving energy from continuous monitoring to few
  points in time
• Objective keeping accuracy as high as possible
• At is the users activity at time t

10
Selective Sampling (cont.)

• Select a set of observation times to maximize
  correct prediction of users activity for times
  when no sampling/classification is made
• Minimize the expected loss

Expected loss over all activity sequences a
Selected observation times for a
Maximum of observations
Minimize uncertainty
Conditional plans

• Sequence of decisions depending on
  observations so far, decides when next
  observation should be made
• Entropy and dynamic programming used to find
  optimal

11
4 Schemes to Select the Conditional Plan

• Uniform Spacing
• Selects observation times at equally spaced
  intervals
• Random Spacing
• B random length observation times selected at
  random
• Exponential Backoff
• Maintains a maximum step size ?max
• If cur. actlast detected act., multiply ?max by
  a else ?max1
• Actual step size ? chosen uniformly at random
  from 1, ?max
• Next observation made at t ?
• Entropy-based
• Minimizing uncertainty using the entropy
  criterion
• Taking transition probabilities of states into
  account
• More frequent sampling for activities with short
  durations

12
Selective Sampling Experiments

• Four new objects performing hour-long activities
• Subjects were indirectly asked to perform
  representative tasks at random times
• User activities manually annotated by an observer
• Resulting in ltactivity,durationgt pairs sampled at
  6Hz
• Data then partitioned into sequences of 5 seconds
• These blocks labeled with annotations and
  classified using pre-trained classifiers in the
  frequency domain

13
Results
Continuous Sampling
Factor of 2 improvement
Classification accuracy lower than previous
experiments due to 1.new subjects 2. noisy
real-world environment
Competitive for low frequencies
Using annotated data as exact classificationNo
SVM (focusing on sampling)
Using classifier output instead of
annotationsErrorSampling Classification
_at_ 6Hz factor of 2.5 improvement
Roughly similar behavior to above experiment
Overall error dominated by classification from
SVM and not by sampling
14
Conclusion

• High efficiency and accuracy for low range
  frequency of 1-10 Hz.
• Competitive classification accuracy for the
  highly erratic and ambiguous (but convenient)
  wrist-based sensing
• Four selective sampling strategies to further
  reduce the resource usage

15
Comments

• Using FFT for each dimension separately looses
  the correlation of among dimensions
• Semi-controlled user behavior for test data
  generation
• Authors assume continuous state change in a close
  set of predefined activities i.e., at any given
  time, one of these activities are taking place