robocrunch
Miles Cranmer
@MilesCranmer
Astro PhD candidate @Princeton trying to accelerate astrophysics with AI. Prev @McGillU @Harvard @DeepMind. ✨🤖🇨🇦
Tweets by Miles Cranmer
In your opinion, what is the most intuitive way to illustrate a latent space? In the past I've simply gone with a long rectangular box, and labeled it as "latent representation"... I feel like there should exist a way that offers a deep learner's intuition about them
Shared by
Miles Cranmer
at
5/10/2022
Awesome @QuantaMagazine @walkingthedot article on interpretable machine learning for science https://t.co/xSCC45wffk Guest appearances by @rogertgn @laurezanna @hodlipson, myself, and others. Discusses methods such as SINDy (@eigensteve++), PySR, Eureqa, NN interpretation w/ SR!
Shared by
Miles Cranmer
at
5/10/2022
Most mind-blowing #Dalle I've seen thus far... Developments in generative machine learning always seem to exceed the *upper bound* of my own expectations
Shared by
Miles Cranmer
at
5/10/2022
PySR now has cluster management capabilities! Scale up symbolic model searches to thousands of cores, directly from python. A nice part about evolution is how parallelizable it is: more cores => more concurrently evolving populations. (Built on ClusterManagers.jl & PyJulia)
Shared by
Miles Cranmer
at
5/8/2022
Have you used PySR or SymbolicRegression.jl in your research to discover or rediscover a symbolic model? If so, consider showing off your paper on the new research showcase page! https://t.co/8QbilrrCXK
Shared by
Miles Cranmer
at
5/3/2022
Symbolic regression efficiently parallelized over *multiple compute nodes*: https://t.co/618dfsDQLm Literally a one-line change with @JuliaLanguage's ClusterManager.jl! Here's 4 compute nodes controlled by 1 Julia REPL(!):
Shared by
Miles Cranmer
at
4/22/2022
e.g., I think if you are training a classification problem, there isn't really any other metric you need to worry about. If you classify data accurately, you are good. But for high-dimensional output where only partial metrics are available, I think Goodhart applies.
Shared by
Miles Cranmer
at
4/22/2022
Is there a mathematical formalization of Goodhart's law in machine learning? "When a measure becomes a target, it ceases to be a good measure" I think Goodhart's law is only applicable when you have an *imperfect* metric (e.g., GDP of an economy or MSE of a fluid prediction).
Shared by
Miles Cranmer
at
4/22/2022
Extremely cool economics paper applying PySR + GNNs to learn symbolic models for international trade! https://t.co/reo47T3ROY By Sergiy Verstyuk and Michael R. Douglas (@HarvardCMSA)
Shared by
Miles Cranmer
at
4/14/2022
This approach allows us to do model discovery even when missing crucial information about our system! I anticipate this being very useful for model discovery in real-world datasets. For more details check out the paper https://t.co/jDeOUt7yFv.
Shared by
Miles Cranmer
at
3/7/2022
Could machine learning rediscover the law of gravitation simply by observing our solar system? With our new approach, the answer is *YES*. Led by: @PabloLemosP With: @Niall_Jeffrey @cosmo_shirley @PeterWBattaglia Paper: https://t.co/jDeOUt7yFv Blog: https://t.co/eP04RXjrYz
Shared by
Miles Cranmer
at
3/7/2022
GitHub Copilot works so well for LaTeX semantic completion that I switched to using VSCode instead of Overleaf's UI. The "LaTeX Workshop" extension is amazing. Live PDF, jump-to-location, preview equations on hover, easier error debugging, etc. https://t.co/AlJklLLJx3
Shared by
Miles Cranmer
at
12/3/2021
This is a cool milestone for PySR. I have been told that Eureqa, the proprietary symbolic regression software, fails to discover simple non-integral power laws from data, such as: y = 10.5 * x ^ 3.1 From 100 samples, PySR can find this in minutes!
Shared by
Miles Cranmer
at
11/19/2021