![Active Learning: KLR
7
• KLR: alternative to SVM with probabilistic outcome
o Natural extension to multi-class problem
o Probabilistic outcome (𝑝𝑖 = 𝒀 = 𝑦𝑖 𝑿 = 𝑥 ,
𝑖 = 1, … , 𝐶)
o Put threshold on some uncertainty measure: e.g.
𝑝 𝐶 − 𝑝 𝐶−1 > 𝑇 [2]](https://image.slidesharecdn.com/activelearningfeatureselectionma-171206135717/85/Active-Learning-for-Acoustic-Classification-AND-Power-Aware-Feature-Selection-for-Audio-Sound-Scene-7-320.jpg)
Active Learning for Acoustic Classification AND Power Aware Feature Selection for Audio/Sound Scene
- 1. Active Learning for Acoustic Classification AND Power Aware Feature Selection for Audio/Sound Scene Mulu W. Adhana /Groep T Campus
- 2. Outline 2 • PhD Topic: Active Learning for Acoustic Scene/Event Classification o Active Learning: Introduction, principles, … o Experimental Results o Current Activities: adaptive, transfer learning, … • Power Aware Feature Selection o Proposed Method o Experiments o Conclusions
- 3. Active Learning: Introduction 3 • Traditional machine learning Techniques (Supervised) o Abundant labeled data o Labeling: labour intensive costly and error prone o Domain expert • Semi-supervised learning • Active learning o Seed information (very small in size) o Sample unlabeled data points for annotation o Involves user to actively annotate data point • Reserved for informative/ambiguous ones
- 5. Active Learning in Practice 5 ? automatically assigned ? Small distance to separating hyperplane Ask the user
- 6. Active Learning: Sampling for annotation 6 • How ambiguous? o Distance from separating line o Probability assigned to the data point o Entropy of the probability assigned to the data point
- 7. Active Learning: KLR 7 • KLR: alternative to SVM with probabilistic outcome o Natural extension to multi-class problem o Probabilistic outcome (𝑝𝑖 = 𝒀 = 𝑦𝑖 𝑿 = 𝑥 , 𝑖 = 1, … , 𝐶) o Put threshold on some uncertainty measure: e.g. 𝑝 𝐶 − 𝑝 𝐶−1 > 𝑇 [2]
- 8. Results: KLR Active Learning 8 NAR dataset: • 21 classes • 431 annotated examples • 10-fold cross- validation • Reduce 84% manual annotation
- 9. Active, Transfer and Incremental Learning 9 Annotated dataset from different domain Small Annotated dataset Source-task • From different env’t • With different sensing devices Knowledge Target-task • Slightly different from the source task
- 10. Power Aware Feature Selection 10 • Wearable Sensor Networks draw energy from battery o Sensing, processing raw data, packet size, o bit-depth while acquiring sound scene/event, … • On-node processing. e.g. movement monitoring o Extracting expensive features => Energy depletion • Select inexpensive and information carrying features
- 11. Proposed Approach 11 • Apply two criteria: Error rate and cost of extracting a feature/subset of features(CPU-time) • Incorporating dependency of feature extraction process k<<n
- 12. Proposed Approach… 12 • Sequential forward selection, wrapper approach: o Let 𝑌 holds already selected features, then 𝑆 = 𝑆 ∪ min 𝑋 𝑖 (𝑒𝑟𝑟𝑜𝑟(𝑆 ∪ 𝑋𝑖) + 𝜆 ∗ 𝐶𝑜𝑠𝑡(𝑆 ∪ 𝑋𝑖)/(|𝑆 ∪ 𝑋𝑖|) where 𝜆 tradeoff , 𝑆 ∪ 𝑋𝑖 −feature cardinality 𝑋𝑖 − random candidate feature o Stopping criterion: add feature 𝑋𝑖 to Y such that 𝐶 𝑌 ∪ 𝑋𝑖 > 𝐶 𝑌
- 13. Feature Extraction: Dependency Graph 13
- 14. Experimental Setup 14 • Feature selection on the NAR dataset recorded by humanoid robot NAO Classes Audio Features • 21 classes (door open, door close, fridge open close, moving chair, …) • 431 examples (Sound Scene) 35 (14-MFCC, 12-GTCC, pitch Cross-correlation, Spectral Rolloff, Spectral Flatness, Spectral Flux, Spectral Kurtosis , Spectral Skewness , Spectral Slope, Spectral Spread and ZCR)
- 15. Results 15
- 16. Questions
Editor's Notes
- #8: [1] Karsmakers, Peter, Kristiaan Pelckmans, and Johan AK Suykens. "Multi-class kernel logistic regression: a fixed-size implementation." Neural Networks, 2007. IJCNN 2007. International Joint Conference on. IEEE, 2007. [2] Adhana, Mulu Weldegebreal, Bart Vanrumste, and Peter Karsmakers. "Active Learning for Audio-based Home Monitoring." Proceedings of Benelearn 2016. No. Epub ahead of print. 2016.