There are certain fields that have been around long enough (example: Physics) and principles are well established that a Ph.D. has an extreme advantage. There are fields that are so new and do *not* have decades worth of principles that a PH.D. has less of an advantage.
-
-
This is perhaps why a person with an experimental science PhD may be more capable than a computer science PhD in data science type jobs. The conventional computer science curriculum does not teach what needs to be taught about how models are created about the physical world.
Prikaži ovu nit -
If your goal is to be an expert in theoretical computer science (CS), then a PhD is important. However, an expert in deep learning requires a different set (but overlapping) of knowledge from that found in CS. An analogy is the difference between GOFAI and connectionism.
Prikaži ovu nit -
The ideas of connectionism are very different from the abstractions that come from GOFAI. To understand these, you have to go elsewhere: physics, biology, ecology, economics, psychology, neuroscience etc. It's not found in mathermatical programming.
Prikaži ovu nit -
To conclude, a deep learning PhD who cannot speak coherently about other fields like biology or physics has questionable foundations. Unfortunately, most CS curriculums don't fit in the time to understand adjacent fields.
Prikaži ovu nit -
It's also very important to realize that Deep Learning requires a ton of plumbing (i.e. technology). Many times, a data scientist doesn't have the requisite understanding of the technology underneath their tools. One should thus never trust a data scientist to build software!
Prikaži ovu nit -
I have both a physics and a computer science background. But when I began to study deep learning, I became aware of my knowledge deficiency. This required me to hit the books hard studying evolution, biology, neuroscience, and psychology.
Prikaži ovu nit -
The first principle that every deep learning researcher should come to grips with is that deep learning is more biological than it is mathematical. To formalize this one should study the complexity science literature.
Prikaži ovu nit -
One hint of why this is true is that Deep Learning architectures are "grown" and not "programmed" (in the classical sense).
Prikaži ovu nit -
Furthermore, if you take the myopic viewpoint of mathematical programming then one could argue that Deep Learning is nothing but credit assignment using gradient descent. That's a very impoverished viewpoint!
Prikaži ovu nit -
The only existing general framework that stitches together ideas from biology and computer science can be found in the complexity sciences. So it is indeed astounding that many DL researcher are ignorant of this field.
Prikaži ovu nit -
DL research speaking from the viewpoint of either Bayesian statistics or mathematical programming simply do not speak in the same conceptual framework as those in the complexity sciences. Their ignorance is revealed by the language that they use to describe DL.
Prikaži ovu nit
Kraj razgovora
Novi razgovor -
Čini se da učitavanje traje već neko vrijeme.
Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.
“A PhD is definitely not required. All that matters is a deep understanding of AI and ability to implement NNs in a way that is actually useful (latter point is what’s truly hard). Don’t care if you finished high school.”