Week 09 · Lecture 2 · PDF with keyframes

Lecture 2 · Sources and credits

Serega artwork is reused from Albert Nasybullin’s Deep Learning for Search course. Scientific plots, interactive scenes and diagrams were prepared for Intro to ML. Constructed signals and illustrative EEG filters are labeled separately from measured SKAB results.

SKAB: Katser and Kozitsin (2020), dataset and testbed, GPL-3.0 licence. Local fonts are distributed under the SIL Open Font License: IBM Plex Sans, JetBrains Mono, Newsreader, Patrick Hand and Pangolin.

Validation-only forecast-quality report.

4. Classification: a label for an observed window

5. Forecasting: predict values that have not arrived

6. Anomaly detection: unusual relative to what?

7. Scores rank; thresholds decide

8. Segmentation: label the timeline, or find its boundaries

9. Imputation: what information may cross the gap?

10. The architecture expresses an assumption

11. A causal TCN grows its field of view

12. Memory and attention answer different questions

13. EEG: time is only one of the axes

14. A classical EEG pipeline already has useful structure

15. A temporal kernel has a frequency response

16. One convolution, computed by hand

17. Temporal first. Spatial second.

18. Separable convolution: refine, then mix

19. Follow one EEG epoch through the network

20. An EEG score is only as honest as its split

22. What changes when we pretrain once?

23. Chronos-2: a modern forecasting model we can read

24. A forecasting task is data plus a visibility contract

25. Scaling keeps units from dominating the network

26. A missing value and a measured zero are different

27. A patch becomes a vector, without value quantization

28. Time attention reads within one variable

29. Group attention reads across related variables

30. The head predicts quantiles for the whole horizon

31. Quantile loss makes over- and underprediction cost different

32. Pretraining teaches transferable forecasting structure

33. A forecast can become comfortable with a fault

34. A detector is a chain of choices

35. The lab question: detect the entire abnormal state

36. Macro and pooled answer different questions

37. Rolling features help this model on validation

38. Compare the complete validation pipelines

39. Useful ranking can fail at the chosen cutoff

40. Averaging helps here; max inherits the saturated tail

41. Compare complete anomaly-detection pipelines

42. Better forecasts do not guarantee better state detection

43. Read the horizon and check the interval

44. Read the score, then read the actual alarm