ml degree should be a mix of heavy-ish mathematics and cs. in most cases they are a cs specialization
for example, at my university ml graduates are expected to excell at writing performant parallelizable code and to mathematically understand the models they are building
What's the use of MS level math in ML? Post-hoc rationalization of ML architectures invented by a good guess? IIRC, the batch-norm paper is a shining example: while the technique is useful, it was invented by intuition, and its post-hoc rationalization - all that impressive looking math - turned out to be bs.
Again, what's the use of MSc+ level math in ML? A good BSc course teaches you in the 1st year linear algebra up to the eigenvalues theory, and diff equations up to fairly involved numerical methods. Just these two subjects are way more than necessary for applied ML.
I'd go even further and claim that ML has no theory on its own: it's a bunch of methods based on simple linear algebra. Unlike a proper math subject that begins with axioms, definitions and theorems, an ML course would have nothing like that. One shining example is convergence of an ML model: there is no theory that would predict convergence, so the only way is ad-hoc attemps to massage the data and hyperparameters.
Comments
ml degree should be a mix of heavy-ish mathematics and cs. in most cases they are a cs specialization
for example, at my university ml graduates are expected to excell at writing performant parallelizable code and to mathematically understand the models they are building
What's the use of MS level math in ML? Post-hoc rationalization of ML architectures invented by a good guess? IIRC, the batch-norm paper is a shining example: while the technique is useful, it was invented by intuition, and its post-hoc rationalization - all that impressive looking math - turned out to be bs.
have you ever read an ml paper? besides, good luck getting ml solutions in high-stakes problems to production without good math to back it
doing ml without knowing math is like doing programming without knowing alg & ds
Again, what's the use of MSc+ level math in ML? A good BSc course teaches you in the 1st year linear algebra up to the eigenvalues theory, and diff equations up to fairly involved numerical methods. Just these two subjects are way more than necessary for applied ML.
I'd go even further and claim that ML has no theory on its own: it's a bunch of methods based on simple linear algebra. Unlike a proper math subject that begins with axioms, definitions and theorems, an ML course would have nothing like that. One shining example is convergence of an ML model: there is no theory that would predict convergence, so the only way is ad-hoc attemps to massage the data and hyperparameters.
i'd go even further and claim you have never picked up an ml textbook, let alone an ml paper