Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Better than Deep Learning: Gradient Boosting Machines (GBM) in R - eRum Conference - Budapest, May 2018

szilard
May 11, 2018
2.5k

Better than Deep Learning: Gradient Boosting Machines (GBM) in R - eRum Conference - Budapest, May 2018

szilard

May 11, 2018
Tweet

More Decks by szilard

Transcript

  1. Better than Deep Learning:
    Gradient Boosting Machines (GBM) in R
    Szilárd Pafka, PhD
    Chief Scientist, Epoch (USA)
    eRum Conference, Budapest
    May 2018

    View full-size slide

  2. Disclaimer:
    I am not representing my employer (Epoch) in this talk
    I cannot confirm nor deny if Epoch is using any of the methods, tools,
    results etc. mentioned in this talk

    View full-size slide

  3. Source: Andrew Ng

    View full-size slide

  4. Source: Andrew Ng

    View full-size slide

  5. Source: Andrew Ng

    View full-size slide

  6. Source: https://twitter.com/iamdevloper/

    View full-size slide

  7. http://www.cs.cornell.edu/~alexn/papers/empirical.icml06.pdf
    http://lowrank.net/nikos/pubs/empirical.pdf

    View full-size slide

  8. http://www.cs.cornell.edu/~alexn/papers/empirical.icml06.pdf
    http://lowrank.net/nikos/pubs/empirical.pdf

    View full-size slide

  9. structured/tabular data: GBM (or RF)
    very small data: LR
    very large sparse data: LR with SGD (+L1/L2)
    images/videos, speech: DL

    View full-size slide

  10. structured/tabular data: GBM (or RF)
    very small data: LR
    very large sparse data: LR with SGD (+L1/L2)
    images/videos, speech: DL
    it depends

    View full-size slide

  11. structured/tabular data: GBM (or RF)
    very small data: LR
    very large sparse data: LR with SGD (+L1/L2)
    images/videos, speech: DL
    it depends / try them all

    View full-size slide

  12. structured/tabular data: GBM (or RF)
    very small data: LR
    very large sparse data: LR with SGD (+L1/L2)
    images/videos, speech: DL
    it depends / try them all / hyperparam tuning

    View full-size slide

  13. structured/tabular data: GBM (or RF)
    very small data: LR
    very large sparse data: LR with SGD (+L1/L2)
    images/videos, speech: DL
    it depends / try them all / hyperparam tuning / ensembles

    View full-size slide

  14. structured/tabular data: GBM (or RF)
    very small data: LR
    very large sparse data: LR with SGD (+L1/L2)
    images/videos, speech: DL
    it depends / try them all / hyperparam tuning / ensembles
    feature engineering

    View full-size slide

  15. structured/tabular data: GBM (or RF)
    very small data: LR
    very large sparse data: LR with SGD (+L1/L2)
    images/videos, speech: DL
    it depends / try them all / hyperparam tuning / ensembles
    feature engineering / other goals e.g. interpretability

    View full-size slide

  16. structured/tabular data: GBM (or RF)
    very small data: LR
    very large sparse data: LR with SGD (+L1/L2)
    images/videos, speech: DL
    it depends / try them all / hyperparam tuning / ensembles
    feature engineering / other goals e.g. interpretability
    the title of this talk was misguided

    View full-size slide

  17. structured/tabular data: GBM (or RF)
    very small data: LR
    very large sparse data: LR with SGD (+L1/L2)
    images/videos, speech: DL
    it depends / try them all / hyperparam tuning / ensembles
    feature engineering / other goals e.g. interpretability
    the title of this talk was misguided
    but so is recently almost every use of the term AI

    View full-size slide

  18. Source: Hastie etal, ESL 2ed

    View full-size slide

  19. Source: Hastie etal, ESL 2ed

    View full-size slide

  20. Source: Hastie etal, ESL 2ed

    View full-size slide

  21. Source: Hastie etal, ESL 2ed

    View full-size slide

  22. http://www.jmlr.org/papers/volume13/bergstra12a/bergstra12a.pdf

    View full-size slide

  23. http://www.argmin.net/2016/06/20/hypertuning/

    View full-size slide

  24. Backup Slides

    View full-size slide

  25. All benchmarks are wrong, but some are useful

    View full-size slide