How to guarantee your machine learning model will fail on first contact with the real world.

Recently I had my PhD thesis rejected. As a failure, I am uniquely positioned to recognize failure. While I am unwaveringly enthusiastic about machine learning, I aim to share my insights into failed machine learning modelling from real-world examples in science and industry. This talk is for you if you have an introductory understanding of machine learning and would like to avoid common pitfalls.

Transcript

1. How to guarantee your machine learning model will fail on first contact with the real world Jesper Dramsch PyData Global 2020
2. 99.8 % Accuracy
3. Chihuahua or Blueberry Muffin? Real-World Applications are rarely 100%
4. Augmentations build better models Yet decrease metrics
5. Ignore the possibility Overfitting and Data Leakage
6. Are you overfitting?
7. Is there Data Leakage? [1]
8. Assume everything is IID
9. I I D Independent and Identically Distributed
10. The real world rarely is independent nor identically distributed.
11. Did you account for class imbalances? [1]
12. Always use Accuracy
13. Imbalanced Metrics [1]
14. Cities Dataset for Semantic Segmentation [1]
15. Losses for Semantic Segmentation [1]
16. Collecting more Data is a Better
17. A Good Data Scientist is Data Critical • CERN throws away most of collected data at 25 GB/s [1] • Geophysical data has to be reprocessed for many different use cases [2] • Someone decides on social taxonomies. ImageNet class “looser / failure” as person. [3] • GPT-2 was trained on Reddit comments. Try and ask it about Earth Science. [4]
18. Strategies That Work (Sometimes) • Multiple Interpreters (Inter-interpreter) • Repeat Interpretations (Intra-interpreter) • Take Responsibility to Change Questionable Taxonomies • Collect Representative Samples
19. Cross-Validation solves Everything
20. Cross Validation to the rescue?
21. Class Imbalances call for Stratification
22. Cross-Validation for Time Series Data
23. Cross Validation for Spatial Data
24. Are you Cross-Validating your data preparation? [6]
25. Even Cross-Validation has its Flaws [5]
26. Absolutely ignore Model Simplicity
27. News Item on AI for Earthquake Aftershock Prediction [8]
28. One Neuron outperforms a Deep Neural Network [9]
29. Can we Crash Test our Machine Learning? []
30. Trust any Counter-Intuitive Results
31. Extraordinary Claims Require Extraordinary Evidence
32. Inferring the face of a person from its speech patterns surely is extraordinary [1]
33. “AI” hiring decisions directly from video [1]
34. Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification - Buolamwini and Gebru [1] Critical Perspectives on Computer Vision - Denton [2] Excavating AI - Crawford and Paglen [3] Tutorial on Fairness Accountability Transparency and Ethics in Computer Vision at CVPR 2020 [4] The Uncanny Valley of ML - Andrews [5] Bias in Facial Recognition [6] Research into Bias in ML
35. Subject Matter Experts often forgot more about a Subject than a Data Scientists has learned during a Project
36. Data can often be explained by many hypotheses. [1]
37. Explainability shows how a Machine Learning Model thinks
38. Post-Hoc Explainability will explain “Why?” even on wrong decisions with 99% [1]
39. Calibration of Classifiers [1]
40. Shap Library for machine learning explainability [1]
41. Interpretability Explainable Forests, Linear Models, RuleFit Explainability SHAP, Partial Dependence Plots, Lime, Feature Importance
42. A Machine Learning Model can outperform your Assumptions and Baseline Data
43. Extracting information and establishing relationships is limiting the machine learning model.
44. Speedround: When your Machine Learning Model isn’t scoring perfectly, you can still Spice Up Your Results
45. It is uncomfortably common to hand-select “good” results [1]
46. It is uncomfortably common to overfit on benchmarks to “sell” a method [1]
47. Committing the “Inverse Crime” [1]
48. Measuring what’s easy to measure rather than meaningful
49. • Use nothing else but accuracy • Under no circumstance spend extensive time on validation • Blindly trust counter-intuitive results because the model converged • Explainability is overrated but has all the answer • Take all these points as gospel Main Take Aways