Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of...
Transcript of Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of...
![Page 1: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/1.jpg)
Darwin Phones: The evolution of Sensing and Inference on Mobile
Phones
GZ06 - Mobile and Cloud Computing
Department of Computer Science University College London
Adam Adamou Tommaso Catuogno
![Page 2: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/2.jpg)
Problem description • Background year 2010:
– Growing number of Smartphones – High number of application using sensor data
to extract information. – Phone sensing individual data not reliable
Collaborative inference concept
![Page 3: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/3.jpg)
Outline
• Description of the Darwin Framework
1. General idea 2. Use Case: Speaker Recognition 3. Performance 4. Examples..
![Page 4: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/4.jpg)
Basic Idea
• Classifier Evolution 1
• Model Pooling 2
• Collaborative Inference 3
![Page 5: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/5.jpg)
Classifier Evolution
Model
Sensor data env 1
• It’s an automated approach for acquiring sensors data and updating models over time.
• It’s designed in order to
achieve high robustness to environment change and settings.
Smartphone
![Page 6: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/6.jpg)
Classifier Evolution
Model
Sensor data env 2
Smartphone
• It’s an automated approach for acquiring sensors data and updating models over time.
• It’s designed in order to
achieve high robustness to environment change and settings.
![Page 7: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/7.jpg)
Classifier Evolution
Model
Sensor data env 3
Smartphone
• It’s an automated approach for acquiring sensors data and updating models over time.
• It’s designed in order to
achieve high robustness to environment change and settings.
![Page 8: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/8.jpg)
General Framework
Can we reuse models that have already been built and possibly
evolved on other phones?
![Page 9: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/9.jpg)
Model Pooling
Share models – From Other devices – From Server
Phone2
Phone3
Server
Phone1
Models
Models
![Page 10: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/10.jpg)
Model Pooling Increase Scalability • Devices
doesn’t have to retrain model that has been received from others
Increase quality • Some device
can never achieve a particular data taken from env
Increase speed • Sharing is
faster then training
![Page 11: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/11.jpg)
Collaborative Inference • Combine the
classification results from multiples phones.
• Increase the quality
and robustness of the results with higher confidence
Inferred 3
Inferred 1 Inferred 2
Sensed Data
Inferred Event Confidence
Sharing Results
Final Inference
Local Inference
Model 1 Model 2 STEP 1
STEP 2
Sm
artphone
![Page 12: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/12.jpg)
Gaussian Mixture Models
• x is a D-dimensional continuous-value data vector • w is the mixture weight • g is the Gaussian function with parameter µ , Σ
Den
sity
Value
![Page 13: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/13.jpg)
Use Case to understand better: Speaker Recognition
Classifier Evolution
Model Pooling
Collaborative Inference
![Page 14: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/14.jpg)
Classifier Evolution
Divided in two phases:
• Initial Training (Application Bootstrap) • 30 seconds samples • Creation of starting model • Creation of the starting Baseline
• Evolution Training • Continuous recorded audio • Baseline audio used to understand if evolve or
not the model with the actual input
![Page 15: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/15.jpg)
Initial training
(Application Bootstrap)
• Silence Suppression • Feature Vector
– Create an MFCC of the sound recorded • Speaker model Computation
– Each speaker has a model Mi that is a GMMi that is represented by a tuple Mi = <µi , Σi , wi>
![Page 16: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/16.jpg)
Initial training
(Application Bootstrap)
Training audio
30 seconds Baseline Audio
Model GMM
Baseline
0 30
![Page 17: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/17.jpg)
Training algorithm
Silence suppression
& voicing filter
MFCCstrain
MFCCsbase
M = < µ , Σ , w > Baseline Audio
![Page 18: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/18.jpg)
Evolution training
(Application at Runtime) Steps: • Obtain Likelihood for a 96ms incoming audio chunk • Compare maximum likelihood estimator result with
BL for model already existing • Recruit new audio as part of training data (if necessary) • Re-train classification model • Compute the new baseline as the original BL plus
the newly acquired data
![Page 19: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/19.jpg)
Evolution training
(Application at Runtime)
MLE
Comparison
Data Recruiting
Model retrain
BL comput
e
MFCC
Maximum Likelihood Estimator
![Page 20: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/20.jpg)
Data Recruiting and model re-training
Comparison between input and the existing BL
Evolution Algorithm
![Page 21: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/21.jpg)
Model Pooling
• Each device send to his neighbors the model evolved in it’s self
• The application could also geo-tag a model and upload it on a backend server.
Any other device that move into this area, can download the model from the server.
![Page 22: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/22.jpg)
Collaborative Inference
Collaborative inference phase can be
broken into three steps: Devices must be synchronized for make
collaborative inference
Local Inference(each phone)
Local inference propagation
Final inference
![Page 23: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/23.jpg)
Local Inference
• Goal: Locally derive confidence level of
the inferred event in order to communicate it to the neighbors
E is true if: 1) Event E is detected at least 1 in the past two previous
windows 2) Confidence accuracy at least 50% larger than any
other event
96 ms chunk MFCC + MLE Window block 3
Window block 2 Window block 1 E is True
E is not true
![Page 24: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/24.jpg)
Local Inference - Algorithm Final result of this phase for phone j: Where
![Page 25: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/25.jpg)
Local Inference propagation
• A timestamp associated with each Lii (Align local inference results I time)
• A Weight also associated with vectors (Determine quality of inference) • Each node propagate the calculated LI
vectors
![Page 26: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/26.jpg)
Final Inference
• Goal: exploit the confidence model calculated by each device, to increase accuracy of solution
where Prob(sJi) is the probability that speaker J is
detected by phone I The independence of Lii facilitate the calculus:
![Page 27: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/27.jpg)
Privacy And Trust
• Raw data never leave the phone • Content of voice recognition is never
disclosed • Users have the option to drop out Darwin.
![Page 28: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/28.jpg)
Experiments
• Experiment: – 8 people – Over two weeks – Recorded audio in different locations – Classifier trained indoors
• 2 phones were used: – The Nokia N97 – The Apple iPhone 3G
![Page 29: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/29.jpg)
The Need For Classifier Evolution
• Why do we need Classifier Evolution? – People having a conversation in an noisy
environment – Voice classifier is trained for a quite indoor
environment – Unreliable results
![Page 30: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/30.jpg)
Experiment One: Walking In A Busy Street
• Three people walking in a busy road having a conversation
• The phones they carry do not have Darwin but have the speaker recognition application.
![Page 31: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/31.jpg)
Experiment One Results: Without Evolution
![Page 32: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/32.jpg)
Experiment One Results: Without Evolution
• Accuracy is low • Trained for quite environments • Not adapting to changes
![Page 33: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/33.jpg)
The Need For Classifier Evolution
• Previous graph shows the need for the classifier evolution.
• Challenge 1 is sensing different environments
• Challenge 2 is separating voice from any other sounds
![Page 34: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/34.jpg)
Experiment Two: Indoor Office Environment
• 8 People in a quite indoor environment • People engaged into a conversation • Phones in different locations and
distances(on table, in bag, in pocket)
![Page 35: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/35.jpg)
Experiment Two: With Evolution
• Speaker 8 using basic recognition application
• Recognition Precision is below 50% • Using an individual mobile phone is
challenging. – Phone could have poor sensing context – i.e. in the pocket or affected by other factors
such as noise mixed with voice.
![Page 36: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/36.jpg)
Experiment Two: Results
![Page 37: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/37.jpg)
Experiment Two: Results
• Speaker 8 speaking in an indoor environment
• Phone could be in his pocket • Away from other people
![Page 38: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/38.jpg)
Experiment Two: Results
• Using Darwin a performance boost is registered.
• This is due to collaborative inference – Where phones work together to achieve a
better classification accuracy • The following graph shows this boost.
![Page 39: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/39.jpg)
Experiment Two: Results With Collaborative Inference
![Page 40: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/40.jpg)
Experiment Two: Results With Collaborative Inference
• Speaker 8 speaking in an indoor environment
• Phones work together achieve higher precision and accuracy
![Page 41: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/41.jpg)
Experiment Two: Results With Collaborative Inference
• From the previous graph you can see the positive impact that Darwin had on the classifier evolution
• Accuracy, precision and recall rise over time
• New audio is obtained for re-training • Performance levels off after time • No more benefit from data • Training stops
![Page 42: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/42.jpg)
Experiment Three: Noisy Indoor Restaurant
• Five people having lunch in a noisy restaurant • Three of the five people are engaged in
conversation • Two of the five phones are placed on the table
– Phone 1 & Phone 2 • Remaining phones are in the peoples pockets • The 4th phone is closest to the 4th speaker
and to a nearby group of people having a conversation
![Page 43: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/43.jpg)
Experiment Three: Without Darwin
![Page 44: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/44.jpg)
Experiment Three: With Darwin
![Page 45: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/45.jpg)
Experiment Three: Results
• Phone 1 & 2 can capture better quality of speech because they are further away from the noise
![Page 46: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/46.jpg)
The Need For Classifier Evolution
• Comparison of the indoor quite experiment and noisy restaurant experiment shows only one difference.
• For the restaurant case 78KB – 100KB of data was sent to the backend servers
• For the indoor quite case 16KB – 70KB of data was sent to the backend servers
• Proves that less data is needed to evolve in a quite environment
![Page 47: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/47.jpg)
Time and Energy Measurements
• The baselines for power usage are determined
• Measurements are performed using the Nokia Energy Profiler tool
• Each measurement was repeated five times
• Measurements were taken for the Nokia N97
![Page 48: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/48.jpg)
The Measurements
![Page 49: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/49.jpg)
Measurements
A: One sec audio sampling, B: Silence Suppression, C: MFCC extraction, D: Voicing, E: Local Inference, F: MFCC transmission to the server, G: Model reception from the server, H: Model transmission to neighbours, I: Local inference broadcast, L: Local inference reception
![Page 50: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/50.jpg)
Prospective Applications
• Virtual square application – Social application for a group of friends
• Place discovery application – Use collaborative inference to determine
location • Friend Tagging application
– Exploit face recognition to tag friends on pictures
![Page 51: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/51.jpg)
Example Place Discovery App
• Two friends are having a conversation in the kitchen.
• Their phones through collaborative inference know who is talking and where they are.
![Page 52: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/52.jpg)
Possible Future Work
• Duty cycling for improved battery life – Better understanding of the cycle can lead to
better battery usage and performance
![Page 53: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/53.jpg)
Conclusion
• The Darwin system can furthermore improve on the standard sensing applications on smartphones
• It combines classifier evolution, model pooling, and collaborative inference
• Darwin is very scalable. It can easily be be implemented in vast frameworks that can be used for sensing applications
![Page 54: Darwin Phones: The evolution of Sensing and Inference on ... · • Recruit new audio as part of training data (if necessary) ... – The Apple iPhone 3G . The Need For Classifier](https://reader036.fdocuments.us/reader036/viewer/2022081617/604b9379f9b3095acf5d3904/html5/thumbnails/54.jpg)
References • Emiliano Miluzzo, Cory T. Cornelius, Ashwin Ramaswamy, Tanzeem
Choudhury, Zhigang Liu, Andrew T. Campbell, "Darwin phones: the evolution of sensing and inference on mobile phones," In Proc. of 8th ACM Conference on Mobile Systems, Applications, and Services (MobiSys), 2010, pp. 5-20.