Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

Oleksandr Honchar "Deep learning for signal processing"

73 views

Published on

Data Science Practice

Published in: Engineering
  • Be the first to comment

  • Be the first to like this

Oleksandr Honchar "Deep learning for signal processing"

  1. 1. DEEP LEARNING for signal processing (or time series) Oleksandr Honchar Mawi Solutions
  2. 2. • AI solutions architect @ Mawi Solutions • Dottore magistrale in mathematics candidate @ UNIVR • Calling myself AI expert in Linkedin
  3. 3. SIGNALS IN THE WILD 🦁
  4. 4. SIGNALS IN THE WILD 🦁
  5. 5. SIGNALS IN THE WILD 🦁
  6. 6. MAWI BAND ♥️
  7. 7. CLASSICAL APPROACHES 🔬 • Classification - time domain analysis
  8. 8. CLASSICAL APPROACHES 🔬 • Classification - time domain analysis - frequency domain analysis
  9. 9. CLASSICAL APPROACHES 🔬 • Classification - time domain analysis - frequency domain analysis - nearest neighbors
  10. 10. CLASSICAL APPROACHES 🔬 • Classification - time domain analysis - frequency domain analysis - nearest neighbors - ML
  11. 11. CLASSICAL APPROACHES 🔬 • Classification - time domain analysis - frequency domain analysis - nearest neighbors - ML • Regression - ARIMA models
  12. 12. CLASSICAL APPROACHES 🔬 • Classification - time domain analysis - frequency domain analysis - nearest neighbors - ML • Regression - ARMA models - smoothing / decomposition
  13. 13. CLASSICAL APPROACHES 🔬 • Classification - time domain analysis - frequency domain analysis - nearest neighbors - ML • Regression - ARMA models - smoothing / decomposition - nonlinear dynamics
  14. 14. CLASSICAL APPROACHES 🔬 • Classification - time domain analysis - frequency domain analysis - nearest neighbors - ML • Regression - ARMA models - smoothing / decomposition - nonlinear dynamics - ML
  15. 15. DEEP LEARNING 🦁 TCE conference, 2014
  16. 16. DEEP LEARNING 🦁 • RNN
  17. 17. DEEP LEARNING 🦁 • RNN
  18. 18. DEEP LEARNING 🦁 • RNN
  19. 19. DEEP LEARNING 🦁 • RNN 1.Theoretical infinite memory 2.Multistep prediction ability 3.Truncated implementation 4.Cant train in parallel 5.Difficult to optimize 6.Slow in inference 7.Doubtful superior performance!!!
  20. 20. DEEP LEARNING 🦁 • RNN • CNN
  21. 21. DEEP LEARNING 🦁 • RNN • CNN
  22. 22. DEEP LEARNING 🦁 • RNN • CNN • RNN + CNN
  23. 23. DEEP LEARNING 🦁 • RNN • CNN • RNN + CNN • Autoregressive CNN
  24. 24. DEEP LEARNING 🦁 • RNN • CNN • RNN + CNN • Transformer
  25. 25. MAWI BAND ♥️
  26. 26. Other successes 🦁 Sales forecasting Wikipedia traffic
  27. 27. DEEP LEARNING 🦁 • RNN • CNN • RNN + CNN • Autoregressive CNN • Other tasks
  28. 28. DEEP LEARNING 🦁 • RNN • CNN • RNN + CNN • Autoregressive CNN • Other tasks
  29. 29. HYBRID SOLUTIONS 🐙
  30. 30. HYBRID SOLUTIONS 🐙
  31. 31. TAKEAWAYS 📚 • Signals are everywhere • Autoregressive CNN > CNN > RNN • Cluster in embedding space • Use GANs not just to generate • Combine DL and classics if you can • It works for NLP, speech and other sequences as well!
  32. 32. TCE conference, 2014
  33. 33. Home reading 1. When Recurrent Models Don't Need To Be Recurrent 2. An Empirical Evaluation of Generic Convolutional and Recurrent Networks for Sequence Modeling 3. DEEP TEMPORAL CLUSTERING: FULLY UNSUPERVISED LEARNING OF TIME-DOMAIN FEATURES 4. REAL-VALUED (MEDICAL) TIME SERIES GENERATION WITH RECURRENT CONDITIONAL GANS 5. Time-series Extreme Event Forecasting with Neural Networks at Uber FB: @rachnogstyle MEDIUM: @alexrachnog

×