How AI Is Changing STEM

[Greg Bernhardt]

We asked our PF Advisors “How do you see the rise in A.I. affecting STEM in the lab, classroom, industry and or in everyday society?”. We got so many great responses we need to split them into parts, here are the first several. This is part 2. Read part 1 here. Enjoy!

gleem
AI: Maybe not quite so ready for prime time but it is coming.
I have no particular expertise with AI but I try and follow the development of AI and robotics routinely by checking for significant advances more or less weekly. Many occupations are ripe to be significantly affected by AI as those that are based on executing standard procedures and do not require delicate physical...

Continue reading...

 

[Auto-Didact]

A perspective from an interdisciplinary machine learning (ML) researcher very much on the ground about the direct utility of AI to the theoretical sciences: in the field and in most of academia, AI is defined as a subset of machine learning. In the contemporary literature, as well as in the media, AI typically refers to the modern deep learning methods, i.e. multi-layered artificial neural networks, but there are certainly other branches. In any case, my perspective is most aligned with that of @jack action i.e. ML should be seen as an extension of statistics, or more accurately statistics is seen as a subset of ML.

The promise of AI research - due to its completely general nature - is not tied to some particular field(s), but functions as a possible benefit for all fields, in exactly the same manner that statistics and mathematics can be and constantly gets applied to all disciplines. In particular, both the hard sciences (i.e. STEM fields) as well as the soft sciences are directly treatable with AI. It goes without saying that the soft sciences are nowhere near as matured in their mathematical characterization as disciplines compared to the hard sciences; indeed, this is why some of these disciplines are regarded or classified as soft sciences in the first place!

STEM disciplines due to the very nature of their subject matter, the historical development of these disciplines and the successful and succinct distillation of relevant specific forms of mathematics with respect to particular problems within these disciplines, tend not to benefit immensely from AI research directly; this is essentially because these disciplines are usually already too mature in their mathematical characterization to be able to easily benefit greatly from what conventional AI research - which is still in a stage of infancy - has to offer, at least if the chosen AI methodology isn't tailored to the degree of mathematical sophistication required.

The benefit of AI research in STEM research can change quite rapidly in the hands of an expert researcher in some STEM discipline who also happens to be an expert in ML, but these people seem to be quite rare. The other option is working together in teams which of course brings with it the same type of differences in culture as between e.g. mathematicians and physicists or between theoreticians and experimentalists, who often essentially speak different dialects and may even have different goals which makes effective communication difficult if not outright impossible.

Generally speaking, the hard sciences have the inherent benefit of tending to have relatively simply delineated core subject matters; it is the sharpness of these boundaries which makes most phenomena studied within the hard sciences relatively easy to be experimentally approached in a straightforward manner. Moreover, this simplicity then tends to be exacerbated by the fact that these experimental findings then also often end up being describable and extrapolatable in a linear time-invariant fashion, i.e. the core mathematical structure captured by most hard science theories that describe various phenomena tend to be intrinsically as simple as possible well, in the sense that there is a simple ideal case which can be easily understood and generalized, mathematically speaking.

On the other hand, most phenomena studied by the soft sciences tend to have neither of the above benefits of inherent simplicity which is characteristic of the phenomena studied by the hard sciences. Moreover, even when a relatively straightforward model in the soft sciences is found and happens to be generalizable, such models then usually tend to be already as laden with complexity as can be, which essentially makes such a simple model already essentially non-ideal. This can all be captured in the following truism: all ideal theories/models are ideal in the same way, while all non-ideal theories/models are non-ideal in their own particular manner.

To summarize, at this stage of infancy of ML, one should not expect a great contribution to STEM fields using AI, precisely due to the relative mathematical maturity of STEM fields in general. This might change within a few years due to the maturing of the field of ML when the dust has settled around which methods to use for solving particular classes of problems. For the most difficult outstanding physics problems - such as Navier-Stokes existence, quantum gravity and open system non-equilibrium statistical mechanics - it is simply too early to say what one should expect.

In contrast, even at this stage of infancy of ML there can relatively easily be a great contribution in soft sciences, precisely due to the relative mathematical immaturity of these soft science disciplines themselves, and AI actually being able to detect simple numerical patterns hidden in the data quite readily that humans simply are unable to recognize using more conventional methods such as statistics. These direct benefits can be so striking that a soft science can actually be transformed into hard science (e.g. think disciplines like mathematical sociology, quantitative finance, biophysics etc).

Before the current ML revolution the only way that such a soft-science-to-hard-science transformation happened in the last century was when some lost mathematician(s) or physicist(s) specifically decided to dabble in some soft science discipline, typically because some readily identifiable equations (at least for physicists) were unrecognized and misunderstood within those disciplines, while also being of direct interest within the mathematician's or physicist's own original discipline (e.g. the equations of Brownian motion being identified in economics by Bachelier a few years before Einstein derived them). In any case, this time will be remembered as the golden age of ML; we are living in quite exciting times.

 

[gleem]

Two interesting things occurred in the news in the last two days.

Musk says that Tesla will have level 5 autonomous driving system by the end of the year. That is a system for a vehicle with no human controls. How long it takes to incorporate it into an actual vehicle or obtain approval to use it is not known.

COVID -19 has put pressure on Tyson Corp. to introduce automation into its chicken the deboning processes. The competitor Pilgrim's Pride has a system that is claimed to yield 98.5% of the meat compared to humans. The difference of 1.5% is equivalent to shutting the operation down for only a week.

COVID -19 is going to be a big driving force to introduce AI/automation as companies evaluate its impact and the impact of possible future pandemics. More stay at home jobs means less and small office facilities for companies driving down the need for facility support staff, reducing commuting congestion reducing pollution. Some good some maybe not. AI will fill in the gaps or take over jobs faster than we might have expected.