3 ms·
> I have a PhD in machine learning Do not take that badly, but have you learned actual statistics and the mathematics behind them? Or like the few ML people I'
by csdvrx 5y ago
> I have a PhD in machine learning
Do not take that badly, but have you learned actual statistics and the mathematics behind them? Or like the few ML people I've met, do you just use and tweak models?
The part you quoted is extremely clear to me, even without actual pictures: it just means the approximation you use departs from the actual function, tell you where and why.
I can infer there will be an artifact, and how I could try to minimize it.
> Sometimes Gibbs phenomenon leads to an artefact (called Gibbs artefact) that we see in our images of the heart when we are doing a certain sequence called stress perfusion cardiac MRI.
You describe where and when it happens in practice, not why it exists in theory.
Different needs for different people!
- mhh__ 5y agoI think you have stated it rather bluntly, but I have noticed this from some people - I don't know whether it applies to the commenter or not, so I won't speculate, but there have been times when - while I don't doubt that they are an expert is the study and ethics of machine learning - I noticed a lack of holistic knowledge when it came to (say) the hardware that actually execute that machine learning.
- csdvrx 5y ago> I think you have stated it rather bluntly, Oops. I was just talking from experience, where I had some expectations of understanding given the level of expertise professed that weren't met in practice. I say that and I don't even have a PhD (or any other degree!!) in statistics or ML... > I noticed a lack of holistic knowledge when it came to (say) the hardware that actually execute that machine learning. Exactly, there are a lot of people I talk to who seem to use "black boxes", without any understand of how it works. If they can be more productive, why not? However, when one of their black boxes is broken, they often get stuck: as they can't even understand how it works in normal cases, understanding when it may stop working, or even more importantly, how to fix it then, is just not possible. I have found that to often be the case with mathematics and statistics. Sometimes it can be a problem, so I have derived some quick test questions to evaluate the person understanding. They are not Shibboleth: they are not hurtful by virtue of having different "levels" of answer that the person may give. Typical example for stats/ML: "Can you explain why the LLN happens?" Bonus points if the person just tells me the Normal distribution is defined by a Gaussian that is the Fourier transform of itself.