industrial machine learning (sigkdd17)
TRANSCRIPT
SIGKDD 2017 (Halifax)Josh Bloom (GE Digital) @profjsb Industrial Machine Learning
Industrial Machine Learning
Josh Bloom VP Data & Analytics, GE Digital Professor, UC Berkeley @profjsb
SIGKDD 2017 (Halifax)
COPYRIGHT 2012-2017, WISE.IO INC.
‣ Brief Background/Introduction
‣ IML Challenges/Opportunities• Optimization metrics • Physics & Data-driven models• Interpretability / Regulation• Privacy & Security
Agenda
SIGKDD 2017 (Halifax)Josh Bloom (GE Digital) @profjsb Industrial Machine Learning
Teaching
‣ Python Bootcamps 200+ undergrad/grad
‣ Python for Data Science graduate course
‣Berkeley Institute for Data Science (BIDS)
Co-founder
Industry
‣ML Applications Company
Code / Repos
Q4’16
CTO, Co-founder Professor, UC Berkeley
Research
Gordon & Betty Moore Foundation
Data-Driven Investigator
‣ Automated Data-driven Discovery & Inference in the Time Domain
‣300+ refereed articles
wise.io
Beyond “Smart” Thermostats, Fitbits, and Self-driving cars…
Industrial Internet of Things: 50b+ connected devices, 600ZB/yr (2020)
SIGKDD 2017 (Halifax)Josh Bloom (GE Digital) @profjsb Industrial Machine Learning
Preventative Maintenance Failure/Anomaly Detection
Assistive Diagnosis & Treatment Systems Optimization
SIGKDD 2017 (Halifax)Josh Bloom (GE Digital) @profjsb Industrial Machine Learning
False + True +
True – False –
Great movie Rec!
Didn’t see an ad I would have hated
!Spam in my
inbox!
Received a box of clothes I
detest
ML ImpactConsumer
False + True +
True – False –
Industrial
Avoided $200k unneeded per engine repair
Identified cancerous
tumor early
Missed a fatal gas pipeline
crack
Took a MegaWatt power
station offline unnecessarily
"
SIGKDD 2017 (Halifax)Josh Bloom (GE Digital) @profjsb Industrial Machine Learning
Implication for the Optimization Metric
‣ Higher accuracy More value
A rational case for Bias in IML
‣ Drive towards higher precision / Low FPR at Recall~1
Credit: Ratish Dalvi
SIGKDD 2017 (Halifax)Josh Bloom (GE Digital) @profjsb Industrial Machine Learning
https://www.slideshare.net/JuneAndrews/counter-intuitive-machine-learning-for-the-industrial-internet-of-things-77853709
Implication for the Optimization Metric
Industrial Setting Results in Choosing the Lower Bound of Positive Training Samples
Image Source: “ROC curve analysis” MedCalc.org
Optimize Cost Function with Infogain Loss Layer (Caffe):
Practical path: Fit noisy FN tail with (skewed normal/Beta), project lower bound
= CTN|TN|+CTP|TP|+CFN|FN|+CFP|FP|
= ε·|TN|+CTP|TP|+∞·|FN|+CFP|FP|
which classifier is best?depends...
SIGKDD 2017 (Halifax)Josh Bloom (GE Digital) @profjsb Industrial Machine Learning
Physical Models Data-driven Models
✓ Captures a priori understanding of device/system. Little data required. ✓ Simulate/anticipate futures not yet seen in the real-world ✓ Parameters have physical meaning: interpretable
Do not learn from new data/outcomes Incomplete physics → imperfect model Refinement is a very manual process
requiring deep domain experts Difficult to incorporate non-physical
metadata
Often requires previous exemplars. Difficult to predict new failure modes
Uncertain generalizability Black boxy: Difficult to drive intuition/
improvements from model parameters.
✓ No explicit physical understanding required ✓ Naturally improves with more data/outcomes ✓Extensible to new data sources
SIGKDD 2017 (Halifax)Josh Bloom (GE Digital) @profjsb Industrial Machine Learning
Marrying Physics & Data Driven Models for IML
‣ Manual featurization - fits to physical models as input ‣ RL (AI gyms/supercomputer simulations) + transfer learning with updates in the real world
‣Modeling residuals of physical models
12 Shah, Dey, Lovett, and Kapoor
(a) Circle maneuver (b) Square maneuver
(c) Space-Time Plot for Circle (d) Space-Time Plot for Square
Fig. 4 Evaluating the differences between the simulated and the real-world flight. In top figures,the purple and the red lines depict the track from simulation and the real-world flights respectively.
factors such as integration errors, vehicle model approximations and mild randomwinds.
Sensor Models: Besides evaluating the entire simulation pipeline we also inves-tigated individual component models, namely the barometer (MEAS MS5611-01BA), the magnetometer (Honeywell HMC5883) and the IMU (InvenSense MPU6000). Note that the simulated GPS model is currently simplistic, thus, we only fo-cus on the three more complex sensor models. For each of the above sensors we usethe manufacture specified datasheets to set the parameters in the sensor models.
• IMU: We measured readings from the accelerometers and gyroscope as thevehicle was stationary and flying. We observed that while the characteristicswere similar when the vehicle was stationary (gyro: simulated variance 2.47e�7rad2/s2, real-world variance 6.71e�7 rad2/s2, accel.: simulated variance 1.78e�4m2/s4, real-world variance 1.93e�4 m2/s4), the observed variance for an in-flight vehicle was much higher than the simulated one (accel.: simulated 1.75e�3m2/s4 vs. real-world 9.46 m2/s4). This is likely in real-world the airframe vi-brates when the motors are running and that phenomenon is not yet modeled inAirSim.
• Barometer: We raised the sensor periodically between two fixed heights: groundlevel and then elevated to 178 cm (both in simulation and real-world). Figure 5(a)shows both the measurements (green is simulated, blue is real-world) and we ob-
Shah+1705.05065
eg., Christiano+1610.03518
‣Impart/impose/imbue physical constraints into models
• Computer vision: e.g., GVNN (s03 layer Euler,…)
• High energy physics: “QCD-Aware Recursive NN for Jet Physics” Louppe+ 1702.00748
Handa+1607.07405
SIGKDD 2017 (Halifax)Josh Bloom (GE Digital) @profjsb Industrial Machine LearningTime Time
Observations
Physical Model Residuals
Final Model Residuals
Physical Model
Model Physical Residuals (eg. LSTM)
pred
ict
t2t3
‣Modeling residuals of physical models
13
Industrial Machine Learning
Applications Still Need to be Understood
“Monitoring the well ‘was a little hard that night’ because so many simultaneous operations were going on…”
Bloomberg - March 13, 2013
SIGKDD 2017 (Halifax)Josh Bloom (GE Digital) @profjsb Industrial Machine Learning
High
LowLow High
Inte
rpre
tabi
lity
Accuracy
Linear/Logistic Regression
Naive Bayes
Decision Trees
SVMs
Bagging
Boosting
Decision Forests
Neural Nets Deep Learning
Nearest Neighbors
Gaussian/Dirichlet
Processes
Splines
* on real-world data setsLasso
Warning
Unscientific &
opinionated!
instance-level feature
importances (ILFIs)
SIGKDD 2017 (Halifax)Josh Bloom (GE Digital) @profjsb Industrial Machine Learning
Anomaly detected last 1.1 hr
Temp 2.2σ higher than last 30 d
Pressure within range
Next Suggested Action • Set field service • Decrease rail speed
s3E-17-ab33Failure in next 30 days: unlikely
Surfacing the WHY, not just the WHAT
Engineer apps to Capture feedback (implicit & explicit) and outcomes → continual learning
SIGKDD 2017 (Halifax)Josh Bloom (GE Digital) @profjsb Industrial Machine Learning
2018: European Union regulations on algorithmic decision-making and a “right to explanation”
Consumer ML Industrial ML
Regulatory Oversite
Massive regulations/bodies (HIPAA, FAA, FDA, NTSB, NRC, …) but
For now, there is no standing/legal contemplation for continual learning models
Bryce Goodman, Seth Flaxman
SIGKDD 2017 (Halifax)Josh Bloom (GE Digital) @profjsb Industrial Machine Learning
Some Paths to Private & Secure IML
• “Privacy-preserving generative deep neural networks support clinical data sharing”
‣Differentially private GANs for as a surrogate for Data sharing
‣Enclave computation / homomorphic encryption
Bost+NDSS15, also Bos 10.1016/j.jbi.2014.04.003• “Machine Learning Classification over Encrypted Data”
• “CryptoNets: Applying Neural Networks to Encrypted Data with High Throughput and Accuracy”Dowlin+MSR-TR-2016-3
‣Privacy Preserving Distributed (Deep) Learning
Beaulieu-Jones+17 https://doi.org/10.1101/159756
• Creating incentives for data sharing across adversaries (cf. Mike Jordan work) • Shokri & Shmatikov CCS15, Abadi+CCS16
SIGKDD 2017 (Halifax)Josh Bloom (GE Digital) @profjsb Industrial Machine Learning
• Secure Domain-Specific Approach for Prediction on and Transmission of Data with Compression and Privacy Preservation
D. Eads (MLConf, sub.)
JSB (62/508,246.)
Some Current Work
github.com/wiseio/paratext
• Data structure to collapsing hyperparameter optimization
• Memory-efficient data frame optimized for ML
• Testing tool to monitor ML workflows given changing data (stability)
SIGKDD 2017 (Halifax)Josh Bloom (GE Digital) @profjsb Industrial Machine Learning
Academic Partnerships
Ray + Plasma Clipper Alegro Arx, …
Weld
New
Founding Partner
SIGKDD 2017 (Halifax)Josh Bloom (GE Digital) @profjsb Industrial Machine Learning
The Power of a 1% Gain in Efficiency
$27B$30B
$63B$66B
$90B
RailAviation
HealthcarePower
Oil & Gas
Source: “Industrial Internet Pushing Boundaries of Minds & Machines” GE, 2012