It's also a lot of fun. For anyone that likes math, but also likes weird approximations, there's all kinds of juicy stuff in scientific computing. It will make you a better ML coder, too.
"Non-linear" threw me off for a second - I almost never see the mathematically correct definition of linear in computer science spaces. For anyone wondering, Celsius to Fahrenheit is an affine transform, technically not linear, because you have to add an offset, not just multiply.
On the other hand, an equation of the form y = a x + b is a linear equation. If you have Celsius and want Fahrenheit you accomplish that by applying a linear equation (F = 1.8 C + 32), so I certainly can't fault people for saying that the transformation they are doing is linear.
I wonder what people would say for something using an equation of the form y = a x^2 + b x + c to transform something? I can't say that I've heard anyone talk of quadratic transformations. On the other hand, I can't think of ever transforming anything with a quadratic equation, so never had the need t speak of it.
(Also, he called it a linear conversion, not a linear transformation).
Trying to predict where electrons go in a molecule, but using a classical model. This can be done with supervised machine learning - you can use quantum mechanics to get lots of labeled data - but it's a tough problem, because chemical physicists have very high standards for accuracy.
I work at Schrodinger Inc, collaborating with the people who make TorchANI (https://aiqm.github.io/torchani/). This currently predicts only molecular energies in the public codebase, but I expect they will add electron density using the same framework.
Currently I am using hybrid DFT (wB97X-D) with plans to move up to wB97M-V and possibly Quantum Monte Carlo. The TorchANI group likes wB97X and up-training with DLPNO-CCSD(T).
Density functional theory is still quantum mechanics, but operating on the expected electron density itself, rather than the many-body problem of all the electrons. It's pretty good, but not fast - around a CPU-minute for a medium-sized molecule.
I'm working on approximating the electron density using just the nuclei positions and some neural networks. The throughput is tens of thousands of times higher than for DFT. But since the model itself contains little or no physics, the training data has to be very clean and complete.
Would you be able to leverage any of the work that's being done around so called "neural ODE's"?
The Julia community seems to be doing some really cool at the intersection of the 2 fields and it seems like it could be useful if you've already got some kind of pre-existing model/structure to hang the ML part off:
Oil deposits are permeable rock with a mix of oil and water in the pores. Pumping oil down into such a formation would be slow and expensive. It would also take different equipment from that which is there at the wellheads now, so you couldn't get started right away.
Existing vaccine efforts will probably succeed, but the possibility that they fail is real and quite scary (to me, anyway). So I'm glad we have a bunch of longer-term speculative projects as backup.
Just because people are out and about doesn't mean we have to go back to the old value of R. In order to make easing lockdown viable, we need to reduce R for people outside, in every way we can afford. Masks for everyone, of the highest quality we can manage without hurting healthcare. Continue distancing. Constant surface cleaning and air filtration. Mass antibody tests, so we know who has already had it. With all of this, and lockdown for critical outbreak areas, we should be able to keep the curve under control.
I agree with everything here except the reference to QE. Quantitative Easing works to control the money supply only because the central bank can quickly mop the money back up by selling the bonds it previously bought. The central bank can't mop up money it sent to people as wages, so if inflation started, they would have no leverage.
The QE money never has been mopped up, nor can it be mopped up quickly. QE has caused asset price inflation. Also, QE is already back in effect, they're just not calling it QE.
By contrast, paying wages to otherwise idle workers isn't going to make much of a dent in the CPI. If it did, there would be a clear correlation between unemployment rate and consumer prices.
There’s no operational difference between selling old Gilts back to the market and the ability to issue new ones. Under expectations they work the same way.
Spending is automatically recovered by taxation of the flow. If the central bank spends on wages, the Treasury will end up running a surplus.
What about a more visual description, like "safety circle"?
When I was a Boy Scout learning to use an axe, they called the danger zone within reach of the axe the "blood circle". A bit overwrought, but it certainly stuck in my mind. I think they say "safety circle" now.
There are a few computational drug discovery companies that might suit you. For example, I work at Schrodinger, whose motto is "Fight Disease With Code."
