Hoe een efficiënte Machine of Deep Learning backend ontwikkelen?
1 like219 views
This document discusses using LSTM neural networks for time series prediction. It provides examples of using LSTM to predict traffic times based on historical minute-by-minute traffic data and evaluating predictions against actual times. The document also discusses data preparation steps like feature engineering and dimensionality reduction needed before using LSTM on different types of time series data like text, images, or numerical values.
Hoe een efficiënte Machine of Deep Learning backend ontwikkelen?
- 2. LSTM for time series prediction
- 3. Wie? • Howest - Kortrijk • Research Lab van MCT • Internet of Things IT Bachelor – Machine Learning & data architectuur sinds 2015 – Deep Learning sinds 2017 • Huidige Tracks: – Smart Tech & AI – AI Engineer
- 4. Eerste generatie Smart Tech & AI Juni 2019
- 5. Postgraduate ‘Applied AI’ 3 semesters – launch January 2019 – different tracks – practice oriented www.howest.be/aiacademy
- 6. Sci-Fi? 6
- 8. Recommendation systems Personalized recommendations Collaborative filtering Real-time notifications Website personalisation
- 9. Conversational AI Tech Support Agents Chatbots Task-Oriented Dialog Agents (TODA) Future: Cross-domain agents & general purpose AI DoNotPay - https://donotpay-search-master.herokuapp.com/
- 10. When not to use machine learning Easy problems Very few samples The effect of “special” variables
- 11. Problems suited for a machine learning approach? Problems without human experts Problems with human experts but very hard to program Problems where it can be programmed, but it’s not cost effective to implement Sufficient amount of data
- 12. Problems suited for a machine learning approach? Problems without human experts Problems with human experts but ve Problems where it can be programme effective to implement Sufficient amount of data
- 13. Problems suited for a machine learning approach? Problems without human experts Problems with human experts but very hard to program Problems where it can be programmed, but it’s not cost effective to implement Sufficient amount of data
- 14. Problems suited for a machine learning approach? Problems without human experts Problems with human experts but very hard to program Problems where it can be programmed, but it’s not cost effective to implement Sufficient amount of data
- 15. Problems suited for a machine learning approach? Problems without human experts Problems with human experts but very hard to program Problems where it can be programmed, but it’s not cost effective to implement Sufficient amount of data
- 16. Data Performance Traditional machine learning Small neural networks Medium neural networks Deep learning Deep learning or machine learning?
- 17. Do you trust a black box?
- 21. Dense neural network Input layer hidden layer hidden layer hidden layer output layer
- 22. Convolutional neural network (CNN)
- 24. CNN: object detection Real-time: YOLO https://pjreddie.com/darknet/yolo/
- 26. Long Short Term Memory (LSTM)
- 27. 27 LSTM: natural language processing
- 28. LSTM: Time Series Prediction
- 29. Data preparation • Feature engineering • Raw data Input – Text – Images – Numerical data • Dimensionality reduction
- 30. Data preparation • Feature engineering • Raw data Input – Text – Images – Numerical data • Dimensionality reduction •Stemming,… TFIDF •Resolution (MNIST 28x28) •RGB from 0 to 1 https://bigsnarf.wordpress.com/2016/11/17/t-sne-attack-data/
- 31. Data preparation • Feature engineering • Raw data Input – Text – Images – Numerical data • Dimensionality reduction • PCA • t-SNE https://bigsnarf.wordpress.com/2016/11/17/t-sne-attack-data/
- 32. Raw data • One CSV file per minute (+40k/month) • 500k lines (BENELUX) • SegmentID; Traveltime in ms
- 33. LSTM http://colah.github.io/posts/2015-08-Understanding-LSTMs/ Long Short Term Memory (SimpleRNN,GRU) Ideal for time series data 8 weight matrices to optimize compared to 1 for Dense
- 34. Model predicts only next minute LSTM one to many
- 35. LSTM one to many residual Model predicts only change for next minute
- 36. Evaluate
- 37. Evaluate Error ±1 std dev
- 38. You can’t buy intelligence
- 39. Successful outcome is not guaranteed
- 40. Bouw uw netwerk uit met bedrijven die op hetzelfde probleem werken Maar geen concurrent zijn
- 41. MCT Research • Machine Learning & Artificial intelligence • Optimaal gebruik maken van Azure/AWS/GCP • Snelheid van AI implementatie & precisie opdrijven • Data Analytics • Meest efficiënte data architectuur • Stresstesting • Wat kan jouw applicatie aan? • 11 Jaar • 80+ bedrijven • 16 Projecten
- 42. The future of exercising (tetra dossier) WIE BOUWT MEE AAN DE FITNESS VAN DE TOEKOMST? MIXED REALITY, AI, SENSOREN EN PRODUCTONTWERP MEER INFO: eefje.battel@howest.be
- 43. LSTM for time series prediction