omrjml

This sounds interesting. Would you mind sharing how exactly you have this setup?

Of course sine can appear in the solution for dynamical systems but the function itself is not dynamical. When evaluating sin(x) you do not need to know about the previous state.

Edit: should not be approximated

Periodic function like sin(x) are not a dynamical system so its previous state does not determine the current state. So it should be approximated in that way.

Not MLP, but PySindy[0] might be a way to find a closed form solution. A simple MLP may not be best way to achieve what are you are after.

[0] https://github.com/dynamicslab/pysindy

This does not appear to be the usual approach of training the neural network on tons of data from the physics simulation. Instead they use the actual physics equations to form the loss function, which is a far more robust way of creating such an emulator. So physics based deep learning title is appropriate.

Modern use for it: https://www.newscientist.com/article/2277794-century-old-wat...

That really resonated with me. I basically stopped using Twitter because of the reasons you outlined here. This tool, however, might make Twitter useful again for me!

From the actual journal article: "Three further points are relevant to the question of whether the model-observed differences in Figs. 10A-C are mainly due to underestimated observed tropospheric temperature trends or to an overestimated satellite WV trend. First, independent estimates of tropospheric temperature change from GPS radio occultation (RO) and radiosondes suggest that over the 2002 to 2018 period of overlap between MSU/AMSU and GPS-RO, tropospheric warming is smaller in microwave sounders than in GPS-RO or radiosondes (Steiner et al. 2020). Second, there is some evidence that observational uncertainties may be smaller in satellite WV data than in satellite tropospheric temperature data (Wentz 2013; see Section 2c). Third, when the individual trend components of our four trend ratios are examined, the agreement between models and observations is better for WV and SST trends than for TMT or TLT trends. These three lines of evidence, taken together with the results of the RSS sensitivity tests, suggest that underestimated observed tropospheric warming is plausible. This inference is predicated on the assumption that the model-based covariance constraints are realistic."

The radiation is not pointed in the same direction as jets like a laser. There is bulk motion of the plasma in a given direction, but the radiation is emitted from particles that are moving in a different directions within that plasma. So the radiation spread is wider than the jet direction. Interestingly, depending on if the jet we are viewing is pointed towards us or not can impact on the radiation signature in the form of blue shift and red shift. If the orientation is just right, bits of jet appear to move faster than to the speed of light, but it's just a geometric effect.

There has been lost of theories to explain how jets from black hole systems form. New observations of light polarization from M87 jet show twisting magnetic field lines which suggest one of the theories might be right. That theory says that spinning black hole's energy can be extracted by these twisting magnetic field lines and therefore power these jets.

That's just one of the jets. In fact, I think only one side is observable as it pointed close to our direction, making the other side very difficult to observe.

Here is a little explanation from TF quantum website. https://www.tensorflow.org/quantum/concepts

Not all vaccines work the same way and it is not obvious what the long term impact of such a "flu pill" would be on our immune system. For example, I would like to know what would be the long term impact on children who do not get exposed to common colds while they are young. Would they go on to develop more autoimmune diseases. They might not, but it is a pretty big unknown.

That's true, I keep forgetting about https://lfortran.org/ Unfortunately in my experience the people I know who almost exclusively code in fortran are not really interested in jupyter notebooks.

It depends on what you are doing. If it is numerical work with matrix operations, then Fortran is fast and pretty intuitive. Some of the annoyances come from dealing with legacy code and mixing it with modern standards.

I am all for vaccines for really dangerous diseases, but we also need to get ill to keep our immune system active and healthy. For example, there are links between being in too clean an environment as a baby and then developing leukaemia. There is similar evidence for other autoimmune diseases.

https://www.sciencedaily.com/releases/2018/05/180521131746.h...

Regarding dynamical systems, Neural ODEs hold promise as they are more analogous to the numerical solvers. I think with any dynamical system, purely data driven approach can be problematic because, as you say, the standard NN architectures are not great for it. You can however add physical constraints, e.g. minimize the Lagrangian as the objective function, to bring stability to these NN emulators.

There are some complex dynamics that take place when mass transfer is happening. In these systems this is via Roche Lobe overflow (1) which depending on the masses of the objects can become unstable after a large mass transfer. But going back to your original point, I guess there may be some configuration that leads to what to you are suggesting. Might be an interesting thing to model to see if it is possible!

(1) https://en.m.wikipedia.org/wiki/Roche_lobe

Yes, they would gain mass from feeding but not get 40% bigger while still in that binary configuration. So it seems these BHs start out bigger than previously assumed.

The problem here are the writers, not the readers. Most people are just not good writers and good at communicating complex ideas clearly. But the obsession with publishing research in journal articles format means there are a lot of reluctant writers writing on unengaging platforms.

Thanks for the pointers there. You are right, that list is even worse. I'm just finding Medium particularly annoying if I search for anything related to machine learning or data science. It is just full of post after post of people regurgitating other basic content. Or maybe that's an indictment on the state of these fields.

I am curious how this would work in the UK. The UK national met service, Met Office, has a free app, but has to show ads (not sure if there is a legal requirement similar to this case). Its biggest competitor is the BBC weather app which is also tax funded and is not allowed to show any ads. Up until a couple of years ago BBC got its data from the Met Office anyway so the two apps were basically showing the same data. However, now the BBC get its weather information from a private weather company. Anti-competition laws and tax-funded organisations is a complicated business!

Blazars are just quasars with the jets pointed towards us. Quasars are basically galaxies with active super massive black hole in its centre. Luckily for us, our galaxy has a quiet black hole (quiet meaning it's not feeding like crazy). This picture might help https://ned.ipac.caltech.edu/level5/March04/Torres/Figures/f...

BL Lac in the picture is a blazar. So same object is classed differently depending on is orientation with respect to us.

I have been considering purchasing one of these, but keep getting put off by the security aspects of cloud based cameras i.e. accidental or otherwise vulnerability that allows others to view the videos.

This is great. Even displays the cloud cover forecast. Love it.

Francois Chollet has hinted at distributed being addressed. https://twitter.com/fchollet/status/1142515054647971840?s=19

Here is a nice article about that https://www.quantamagazine.org/in-mysterious-pattern-math-an...

It may be possible, with a lot of work, to get the model up and running, but to get any type of accuracy out of the model, realtime data assimilation of observations is key and that is possibly even more complex. So I'll agree with the comment that it may be easier to just hook into a national weather service API. Of course, if you just want to learn about getting these models up and running then it's worth trying. (I work at a national met service developing these models.)