4 ms·
It feels dissatisfying because most of those theories are overfitting of models with tons of free parameters to all available experimental data, and when new ex
by sesm 3y ago
It feels dissatisfying because most of those theories are overfitting of models with tons of free parameters to all available experimental data, and when new experimental data appears a new free parameter is introduced to the model. The actually useful theories, that reduce the number of free parameters are yet to be discovered.
- nologic01 3y agoNot the case. We are talking physical theory not LLM's. It is only in cosmology where there is arguably quite a bit of parametric freedom.
- PartiallyTyped 3y agoAdding, physical theories are predictive. The theory is constructed first and is more constrained than what it aims to predict. It’s not machine learning where you learn a model from data, we build the model and then verify its predictions. We is used colloquially. I am not a physicist. I am a mathematician working as an SDE.
- spacebanana7 3y agoParameter estimation on its own isn't real science. Hypothesis should be statements about reality, like that a given material is a superconductor. With enough lawyering, one can argue that statements around whether model #53sa7r describes a phenomenon accurately qualify as hypothesis. But the success or failure of such hypothesis rarely make for interesting papers to anyone but statisticians. And they don't often contribute to the human ability to reason about reality, build new things or predict the future. Like econometric models that never forecast with accuracy >10 years in advance.
- Nifty3929 3y agoThis exactly. Or at least it seems this way to me, as a non expert with some knowledge. Add another particle, or term, or whatever to make it match. Feels very inelegant. I like your description of overfit, which hadn’t occurred to me before, ironically though I spend a lot of time with ML. When teaching ML I talk about adding loops to the Ptolemaic planets-orbit-the-Earth-model - and how it actually worked pretty well because they spent a lot of time adding loops to the orbits to make it fit. It was overfit.
- occamrazor 3y agoPhysicists haven’t added new particles in a long time.
- Timon3 3y agoAs far as I understand the Standard Model is not just corrected for inaccuracies, it has also been used to make a number of predictions. This means there is some value in it and it's not pure overfitting.
- judofyr 3y agoGoogling "overfitting" gives this definition: > Overfitting is an undesirable machine learning behavior that occurs when the machine learning model gives accurate predictions for training data but not for new data. However, most of the particles introduced in physics have been done with concrete predictions. For instance, when the top/bottom quarks were first proposed (1973-1975) the theoretical paper also included a bunch of predictions. In the following years all of the predictions turns out to be true. Trust me, everyone wishes the Standard Model was simpler and was more "fundamental", but no one has figured out how to make it work. I'm not quite sure how "overfitting" fits here.
- cyanydeez 3y agoSo, basically, AI
- empath-nirvana 3y agoI think it's sort of remarkable how _few_ free parameters there are to describe, like, _all of existence_. How many are there, a few dozen? It's pretty incredible if you think about it.