Distributed TensorFlow and Classification of Time Series Data Using Neural Networks

by Sophia TurolAugust 8, 2016
Being quite a standard problem, classification of time series data can become challenging.

munich_tensorflow_meetup-v11

While the conventional methods may leg behind in terms of accuracy, convolutional neural networks come to a rescue. At the recent TensorFlow meetup in Munich, the speakers highlighted the potential of deep learning tools for classifying time series data, as well as the perks of training models using distributed TensorFlow.

 

Classifying time series data with TensorFlow

In his session, Andreas Pawlik of NorCom explored the potential of deep learning technologies and TensorFlow for classification of time series data, or timestamps used for sequence of events. He enumerated a few examples for applying this approach:

  • Light curves in astrophysics
  • Different shapes (skull, blood cell, butterfly, etc.)
  • Electrocardiograms in medicine
  • Prootein sequences in genetics
  • Intruder activity logs in IT security
  • Sensor data

Andreas underlined that though classification of time series data is a standard problem, still it can be challenging. Furthermore, conventional approaches may be computationally expensive and their accuracy heavily depends on the quality of user input. In contrast to that, deep learning technologies—such as multi-scale convolutional networks—pose a better alternative with high accuracy rate.

Find Andreas’s presentation below.


 

Intro to distributed TensorFlow

Miha Pelko of NorCom overviewed the distributed version of TensorFlow, which allows for training bigger models faster. He demonstrated different ways of model training parallelization for:

  • Separate models with different data
  • Separate models with the same data
  • Full model with split data
  • Split model with full-split data

Recently, we have conducted a study exploring the performance of the distributed TensorFlow. You can find full Miha’s slides below.

 

Blog - Predictive Analytics

Want details? Watch the video!

 

Join the meetup group to get informed about the upcoming events.

 

Further reading

 

About the experts

Andreas Pawlik is a data scientist with a focus on the automotive industry. He discovered his interest in delivering data-driven solutions as researcher in astrophysics, where he developed and used supercomputer simulations to investigate the formation of galaxies.

Miha Pelko is a data scientist for the automotive industry. Formerly, he was a researcher in the field of computational neuroscience with special affinity to deep neural networks.

TensorFlow Cheat Sheet
  •  
  •  
  •