This is a "classic" machine vision task that has traditionally been solved with non-learning algorithms. (That in part enabled the large volume, zero defect productions in electronics we have today.) There are several off-the-shelf commercial MV tools for that.
Deep Learning-based methods will absolutely have a place in this in the future, but today's machines are usually classic methods. Advantages are that the hardware is much cheaper and requires less electric and thermal management. This changes these days with cheaper NPUs, but with machine lifetimes measured in decades, it will take a while.
Way late response: the off the shelf stuff is very very expensive as one would expect for industrial solutions. I was tasked to build something from scratch (our own solution). It was quite the journey and was not successful. If anyone has pointers or tips in this department I would truly love to hear about them!
My initial thought on hearing about this was it being used for learning. It would be cool to be able to talk to an LLM about how a circuit works, what the different components are, etc.
I triggered some bug by pausing the simulation, setting the mass of one of the objects to 29.1, then resuming. The lighter objects bounced into the massive objects a few times, then all three objects were suddenly ejected with a very high velocity.
If you fixed something in an open source library you use, and you don't push that upstream, you are bound to re-apply that patch with every library update you do. And today's compliance rules require you to essentially keep all libraries up to date all the time, or your CVE scanners will light up. So fixing this upstream in the original project has a measurable impact on your "time spent on compliance and updates KPI".
This touches on what I ended up telling them: maintaining a local patchset is expensive and fragile. Running customized versions of things is a self-inflicted compliance problem.
> It distracts from what is actually helping which is using different functions with nicer behaving gradients, e.g., the Huber loss instead of quadratic.
Fully agree. It's not the "fault" of Backprop. It does what you tell it to do, find the direction in which your loss is reduced the most. If the first layers get no signal because the gradient vanishes, then the reason is your network layout: Very small modifications in the initial layers would lead to very large modifications in the final layers (essentially an unstable computation), so gradient descend simply cannot move that fast.
Instead, it's a vital signal for debugging your network. Inspecting things like gradient magnitudes per layer shows you might have vanishing or exploding gradients. And that has lead to great inventions how to deal with that, such as residual networks and a whole class of normalization methods (such as batch normalization).
Since their energy density is still lower, it will probably take a while for them to be adapted in EVs.
But their impact on energy storage to stabilize the grid, both technically and in terms of prices, can not be overstated. Cheap, safe storage is the key component missing in Europe for using more renewables. Without that you need to keep gas plants in reserve, should there be a few days without sun and wind.
I bet we (well, China, at least) will see some lower range but cheaper EVs using sodium batteries pretty much right away. A lot of people would be fine with having something that can only do 100 miles as their daily commute vehicle as long as it was cheap, especially in 2 vehicle families.
The bigger picture is though that taking swings every day also requires a bit of "luck". Some people just don't have the energy, education, mindset, or environment.
It's easy to assume it would be in our control, but if you're just tired all day every day because, say, your hormone balance is off and no one can tell you why, you might statistically accomplish less than others.
> No cases of invasive cancer were recorded in women immunized at 12 or 13 years of age irrespective of the number of doses.
> Women vaccinated at 14 to 22 years of age and given 3 doses of the bivalent vaccine showed a significant reduction in incidence compared with all unvaccinated women
For the second group, cases dropped from 8.4 to 3.2 per 100k.
Deep Learning-based methods will absolutely have a place in this in the future, but today's machines are usually classic methods. Advantages are that the hardware is much cheaper and requires less electric and thermal management. This changes these days with cheaper NPUs, but with machine lifetimes measured in decades, it will take a while.