Pro Yearly is on sale from $80 to $50! »

Активное обучение в машинном обучении

Активное обучение в машинном обучении

Евгений Цымбалов
Data Scientist, Skoltech

14 мая 2020
Online Data Science Meetup

Евгений Цымбалов, исследователь в области анализа данных из Сколтех, расскажет нам об активном обучении в машинном обучении.
В докладе будут рассмотрены основные сценарии, применяемые на практике, подходы на основе оценки неопределенности, ансамблевые подходы. Особое внимание будет уделено рассмотрению алгоритмов для оценки неопределенности на основе нейронных сетей.
Доклад будет вам полезен, если вы интересуетесь анализом данных и разработкой систем машинного обучения.


Подписывайтесь на Exactpro в социальных сетях:


Подписывайтесь на YouTube канал Exactpro



May 14, 2020


  1. Active Learning for Machine Learning Evgenii Tsymbalov, Skoltech Data Science

    Meetup, 2020-05-14
  2. Outline • Introduction. Active learning types • Uncertainty estimation. Ensembles

    • Fast approaches for neural networks • Problems and best practices 2
  3. Passive learning Active learning 3

  4. Active learning Scenarios: • Membership Query Synthesis • Pool-Based Sampling

    • Stream-Based Selection • In many applications, labelled (annotated) data is very limited • Unlabelled data is usually widely available • Labelling (annotation) is often expensive • Thus, a clever choice of points to annotate is needed • Active learning uses machine learning model to select the points for annotation • Applications: industrial design, chemoinformatics, material design, human annotation (NLP, images…) 4
  5. Approaches to active learning • Querying from diverse regions •

    Query by committee • Uncertainty sampling • Variance reduction • Expected model change • Expected error reduction • … easy breezy need a PhD difficulty 5
  6. Query by committee (linear) model that distinguishes red and blue

    points labeled data unlabeled data How to enhance: • more models • more diversity • train on different data • calibrate results 6
  7. Uncertainty estimation EXPECTATIONS REALITY Main assumptions: • model error correlates

    with UE • the most erroneous data is the best for learning 7
  8. Uncertainty Estimation for neural networks State of the art: •

    Classification: BALD (works with ensembles too!) • Regression: dropout with bells and whistles UEs for NNs: • Ensembling (accurate yet costly) • Bayesian NNs (natural but might be complicated to achieve state of the art) • Dropout-based 8
  9. Examples 9

  10. Problems, an interlude Random sampling is actually hard to beat.

  11. Problems, part 1 11

  12. Problems, part 2 Right: oracle is overrated Bottom: warm start

    affects generalization 12
  13. How-to • Use ensembles if you can (SotA, NVidia recommends)

    • NNs: • For classification, use BALD (Bayesian Learning with Disagreement) • For regression, start with MC dropout and then improvise • Other models: tree ensembles also may help • If you know your data or machine learning, look at Bayesian approach • It has a built-in variance (uncertainty) estimation • Always check with the baselines 13
  14. References • Gal, Yarin. "Uncertainty in deep learning." University of

    Cambridge 1 (2016): 3. (and follow-up works). • Settles, Burr. Active learning literature survey. University of Wisconsin-Madison Department of Computer Sciences, 2009. • Ash, Jordan T., and Ryan P. Adams. "On the Difficulty of Warm-Starting Neural Network Training." arXiv preprint arXiv:1910.08475 (2019).Our papers • Tsymbalov, Evgenii, Maxim Panov, and Alexander Shapeev. "Dropout-based active learning for regression." International Conference on Analysis of Images, Social Networks and Texts. Springer, Cham, 2018. • Tsymbalov, Evgenii, et al. "Deeper connections between neural networks and Gaussian processes speed- up active learning." arXiv preprint arXiv:1902.10350 (2019). • Tsymbalov, Evgenii, Kirill Fedyanin, and Maxim Panov. "Dropout Strikes Back: Improved Uncertainty Estimation via Diversity Sampled Implicit Ensembles." arXiv preprint arXiv:2003.03274 (2020). • Rasmussen, Carl Edward. "Gaussian processes in machine learning." Summer School on Machine Learning. Springer, Berlin, Heidelberg, 2003. • Burnaev, Evgeny, and Maxim Panov. "Adaptive design of experiments based on gaussian processes." International Symposium on Statistical Learning and Data Sciences. Springer, Cham, 2015. 14