Papers of the day   All papers

One neuron is more informative than a deep neural network for aftershock pattern forecasting

Comments

Jun 25 2019 KordingLab

unknown to most outside the field, DL often does not outperform far simpler models. With obvious exceptions for NLP, vision, and RL.
9 replies, 223 likes


Jun 25 2019 Thomas Yeo

β€œThis suggests that DL - in fact - does not improve prediction compared to simpler baseline models.β€œ - controversy about deep learning paper predicting earthquake published in Nature. https://arxiv.org/abs/1904.01983
5 replies, 172 likes


Oct 03 2019 π™Άπšžπš’πš•πš•πšŠπšžπš–πšŽ π™³πšžπš–πšŠπšœ πŸ€–πŸ”πŸ§ 

β€œwe reformulate the 2017 results using two-parameter logistic regression (that is, one neuron) and obtain the same performance as that of the 13,451-parameter DNN” πŸ˜†πŸ˜‚ https://www.nature.com/articles/s41586-019-1582-8 #AIhype
2 replies, 104 likes


Jun 25 2019 /MachineLearning

Misuse of Deep Learning in Nature Journal’s Earthquake Aftershock Paper https://www.reddit.com/r/MachineLearning/comments/c4ylga/d_misuse_of_deep_learning_in_nature_journals/
2 replies, 53 likes


Jun 30 2019 F Rodriguez-Sanchez

That's the problem with hype. I'm all excited to see what #MachineLearning will do for us in next years, but not everything has to go Deep Learning now. Quite often simpler models will perform similarly well (or better). See also discussion at https://towardsdatascience.com/stand-up-for-best-practices-8a8433d3e0e8
0 replies, 21 likes


Jul 01 2019 Lars Juhl Jensen

Ouch! Double ouch! "We reformulate the 2017 results in probabilistic terms using logistic regression (i.e., one neural network node) and obtain AUC = 0.85 using 2 free parameters versus the 13,451 parameters used by DeVries et al. (2018)."
2 replies, 20 likes


Jun 27 2019 Alex Yarosh

Arnaud Mignan and Marco Broccardo are out there doing the important work https://arxiv.org/abs/1904.01983 https://t.co/k5appXLZau
3 replies, 12 likes


Jun 26 2019 Jon Huang

Machine Learning burn: AUC from logistic regression (single layer) with 2 parameters ties deep learning neural net with 13k+ features, and BEATS it with 4 parameters. Read about it below the fold.
0 replies, 9 likes


Jun 28 2019 Tom Fawcett

Case study of unfortunate evaluation methodology. Yes, deep learning predicts earthquake aftershock patterns (https://blog.google/technology/ai/forecasting-earthquake-aftershock-locations-ai-assisted-science/) --- but no better than logistic regression (https://arxiv.org/abs/1904.01983).
0 replies, 8 likes


Jun 27 2019 Slawek Smyl

One neuron is more informative than a deep neural network for aftershock pattern forecasting https://arxiv.org/abs/1904.01983 :-)
0 replies, 8 likes


Jun 27 2019 Anthony Lomax 🌍πŸ‡ͺπŸ‡Ί

@InSeismoland @wulwife Opinions on all this from seismologists with experience/interest in ML??? Also see: https://arxiv.org/abs/1904.01983 https://link.springer.com/chapter/10.1007/978-3-030-20521-8_1
1 replies, 4 likes


Jun 25 2019 Gabriel Weymouth

I certainly hope @nature publishes this "unsexy" two parameter model that is just as effective as deep learning...
1 replies, 3 likes


Jun 27 2019 Lim Zhan Wei

One neuron is more informative than a deep neural network for aftershock pattern forecasting" https://arxiv.org/abs/1904.01983 -- a great read. Counter-example for deep learning is great for everything. Also good reminder solution looks simple only on hindsight, a classic Colombus' egg.
0 replies, 3 likes


Jun 26 2019 Giovanni Petrantoni

One neuron is more informative than a deep neural network for aftershock pattern forecasting. "Therefore, the objective of our study is not to restrain the use of DL in this field, but to stimulate a further research effort." This applies to many fields! https://arxiv.org/abs/1904.01983
0 replies, 2 likes


Jul 08 2019 Grady Booch

@tdietterich oh, crap. i forgot the link https://arxiv.org/abs/1904.01983
0 replies, 1 likes


Jun 27 2019 Tim

I couldn't resist https://arxiv.org/abs/1904.01983 https://t.co/Evl9I8Gupt
0 replies, 1 likes


Jun 29 2019 Michael Gamer

Interesting! Thanks for sharing
0 replies, 1 likes


Content