Please register any effective altruism related experiments here in order to avoid publication bias.

Facebook welcome v1 Edit

Date registered: 2014-12-10. Owner: Giles Edkins

Greet people when they join the effective altruists facebook group, using a standard greeting. Measure their participation 3 months after joining according to the following criteria:

  • Donating
  • Joining the EA Donation Registry
  • Creating an EA Profile
  • Running a birthday/Xmas fundraiser (if dates allow)
  • Signing the GWWC pledge
  • Number of EA posts on personal blog or EA forum

Christmas Fundraisers Edit

Record your name, the date the fundraiser started, the target amount and how confident you feel that it will be a success (1-10).

Name: Giles Edkins. Date: 2014-12-10. Target: $200. Confidence: 3.

Neural Network Earthquake Predictor (Round 2) Edit

Name: Joseph Chu. Date Registered: 2014-12-12. Target: Success! Confidence: 1.

A "moon-shot" project to attempt to apply the machine learning technique of artificial neural networks to the problem of predicting earthquakes using foreshock data.

Purpose and Proposal Edit

Among all known natural disasters, earthquakes, and the tsunamis associated with earthquakes, tend to be some of the most destructive in terms of lives lost. Unlike many other disasters however, part of this destructiveness comes from the apparent unpredictability of said events. If it were possible to predict a large magnitude earthquake even just 24 hours beforehand, it is possible that a significant percentage of deaths could be avoided through evacuation. As such, an experiment is proposed to apply state of the art machine learning techniques to try to predict earthquakes using foreshock data.

Dataset Edit

Foreshock data, consisting of earthquakes in a certain window before the earthquake to be predicted, is readily available from the USGS website. Though the data from 1900 to 1973 is exceedingly sparse, the data from 1973 onward is increasingly detailed.

Methods and Material Edit

While a prior experiment using a very basic artificial neural network was unsuccessful, it occurs that the algorithm utilized (a modified Autoencoder/Perceptron hybrid) was not considered state of the art, and this may have contributed to the failure. As such it is proposed that state of the art algorithms such as Convolutional Neural Networks and Deep Belief Networks be applied to the problem.

The experimenter currently owns two high-end computers with fairly high-end Nvidia GPUs that are CUDA compatible and therefore able to implement the Theano library on. Code for this experiment will most likely be in Python, to make use of the Theano library to allow for GPU-based acceleration of the artificial neural networks.

Expectations Edit

It should be clear that the probability of this experiment actually succeeding in its current form is still quite low. The reason why to attempt it anyway is simply that the expected utility is still high because the potential utility of success is extraordinarily high. If it works, we'll have something that could save thousands of lives annually. If it doesn't work, the experimenter will have at least been able to practice some coding and state of the art machine learning techniques.