4. Classification for Time Series
• Time Series: Sequence of Events
• Examples:
– Light Curves in Astrophysics
– Skull (blood cell, butterfly, …) Shapes
– Electrocardiograms in Medicine
– Protein Sequences in Genetics
– Intruder Activity Logs in IT Security
– Sensor data in Industry
– …
• Use cases: Indexing, anomaly detection, …
Keogh et al (2006),
http://www.cs.ucr.edu/~eamonn/shape/shape.htm
6. Conventional classification techniques
(for a review see, e.g., Bagnall et al., arXiv:1602.01711)
• Distance based
– kNN search + majority vote
– Euclidean Distance,
Dynamic Time Warping
– Often difficult to beat in
accuracy
– Computationally expensive
• Feature based
– Machine Learning (logistic
regression, SVM, …)
– Usually less expensive
– Accuracy depends on
quality of features
(distributional properties,
spectral coefficients, …)
– Features are handcrafted
10. Why TensorFlow
“In short, Theano invented the genre. It's the Ford
motors of compiling code for deep learning. But
TensorFlow appears well on its way to emerging as
the Tesla motors of the genre.”
-- Zachary Chase Lipton, UCSD
(http://www.kdnuggets.com/2015/12/tensor-flow-terrific-deep-learning-library.html)
• Python
prototyping
• Distributed
learning using
multiple devices
• Google (where
things move fast)
• Similar to Theano
(yes, it’s a plus)