wise.io: a machine-learning platform (pydata sv 2013)
DESCRIPTION
Video can be found here: https://vimeo.com/63271828TRANSCRIPT
Henrik Brink, CTO
PyData SV 2013, March 19
Machine Learning, Python and Raspberry Pi’s
@brinkar @wiseio
11kly
11kx
Reference New Difference
Palomar Transient Factory (PTF) ~1.5 million candidates per night
~10 new transients
http://www.astro.caltech.edu/ptf/
PTF11kly (SN 2011fe)
©Peter Nugent
Supernova Discovery in the Pinwheel Galaxy
11 hr after explosion
Nearest SN Ia in >3 decades
Python Bootcamps at Berkeley
Machine Learning in
ML pain points I
• Data is messy!
• Hard to scale non-linear algorithms to large datasets
• Ad-hoc feature engineering
• Collaboration on data, features and models is difficult
data scientists
Random Forest®Random Forest® is a registered trademark of Salford Systems
Brink+2012
Brink+2012
10% label noise
bit.ly/YZZb9d
WiseRF™
WiseRF™
12 GB MNIST in 90 seconds
8 core EC2 instance
WiseRF™
A brain for the Internet of Things!
Higher speeds
Larger datasets
Hadoop / Mahout
Energy sensorsAd bidding
Credit card fraud detection
Video tracking
Financial predictions
Batch product recommendations Real-time
recommendations
Internet of ThingsHealthcaresensors
Fast vs Scalable
High-frequency data science
High-frequency prediction
Feature engineering
Data ingestion
Raw data
Model deployment
Model validation
High-frequency Machine Learning
• Statistical validation of models
• Lack of feature engineering expertise
• Dealing with data and computing infrastructure
ML pain points IIapplication developers
Machine Learning as a Service™
bit.ly/15eWEZG
Reusable features
Collaboration
Integration
docs.wise.io
• Scalable infrastructure required
• Hard to go from data science experiments to production.
• Complete privacy / security.
ML pain points IIIbusiness and enterprise