aoki

a meta-comment:

it's weird - i submitted this a couple of days ago (https://news.ycombinator.com/item?id=17018883), but it still doesn't show up in the default HN search view. (we even edited the blog headline the same way!) but it does show up in the "by date" view.

it's not the point of the article, but i find it fascinating that this industry has managed to figure out how to capture >400k wild animals per year and return ~90% of them alive to the ocean after harvesting their blood with a giant needle.

there must be an interesting story here about how they figured out that they needed to do this with a sustainable take rate (unlike most commercial fishing industries).

i am not an RF expert, but i find it interesting that folks would use a custom waveform (presumably, since they're using USRP SDRs) for such applications.

agreed. in the valley you do hear it attributed to intel.

the other long-time intel—ism that has now been appropriated everywhere is the OKR system. it was really weird moving from intel to google and hearing someone from peopleops talk about “this part of google culture called OKRs.”

it appears from linkedin that he was an AF intelligence officer.

being a military intelligence officer is not the same as, say, being a clandestine service officer. apparently, the AF even sees fit to give you "intelligence wings" to wear on your chest. there appear to be an awful lot of current and former intelligence officers who list their specialty and former duty stations in linkedin.

i think it's hard to get kids into these topics (beyond shallow youtube factoid videos) at an age when they're mostly into fun. for the STE part of STEM, you can engage more of them with "maker"-type projects aimed at creative enjoyment rather than at strict utility.

i would find it hard to get #1 Son to do physics experiments for fun on his own, but he's pretty happy to learn about pendulums and gears using tinker crates [0]. somehow it's more interesting when the gears are making a bird mobile flap its wings.

[0] https://www.kiwico.com/tinker

most serious US math programs have honors sequences that are aimed at students who intend to go for a PhD and who are often starting with post-calc-I/II/III courses from the very beginning. tao says right in the introduction that this text is for such an honors sequence at UCLA. so, yeah, this book is explicitly written for undergrads who are being groomed to enter elite graduate programs because they did the work (before college) to fit that profile.

in the US, i believe the real analysis course for non-honors courses at R1s is often based on something like ross, abbott, or bartle & sherbert, whereas for non-R1s (where most math majors will be teachers) it may be based on something like lay or wade instead. these books are more accessible to students with less mathematical maturity.

i think what the reviewer is saying is that it would be a mistake to use this book in one of those non-honors courses with a poor faculty-student ratio. even if you do have some students who have the interest and ambition, you'd be doing a disservice to the rest. you'd be exceeding the level of interest for most and unable to support anybody adequately.

i'm not sure i agree.

google's cloud couldn't host this customer's production application because (afaik) google can't host classified data.

so then you're talking about the fact that your tech is inherently "dual-use." but that's almost anything in computing, absent a license that prevents it. in the early days of postgres, we were a little weirded out by some of the support requests we got [0], given the university prototype nature of the system.

[0] https://www.paulaoki.com/.admin/pgapps.html

indeed - i'm not clear on the basis for the outrage here.

from what the article here says (i don't know more about it than that), it's not clear to me that google is doing more than selling their cloud video analysis products to a customer for experimental use on unclassified data. and this kind of sale necessarily comes with offers of (paid) support and consulting services because customers need help. one might object to this customer, certainly - but if so, one might ask, to which customers is google's cloud business not supposed to sell services on this basis? google has never pretended that it didn't sell its products to the defense and intelligence agencies at all; keyhole (google earth) was for many years a major product in the GEOINT sector [0].

if one's objection is to folks working on anything relating to a military or intelligence problem, well, a very large fraction of US senior faculty have taken funding from defense agencies to work on unclassified versions of applications: tracking, image/video/text analysis challenges, etc. etc. etc.

[0] http://trajectorymagazine.com/genesis-google-earth/

my only goal in mentioning the statistics sequence at all was to give a familiar example where the standard sequences vary in depth depending on audience. a trivial point, yes, but i wanted to be concrete because it's the internet. apparently that was a terrible choice, as it was far too close to the topic at hand; my apologies for making you search so hard for a connection.

i made my request because it's jarring for me as a reader when you punctuate your (often delightful) expository writing with conclusions about entire fields and large organizations that seem (on the face of it) to be justified by old and/or very limited data.

but that's a selfish request, and you are of course free to tell me to get lost and post whatever you want (and i'll still read it); i'm certainly not going to pursue this further, aside from the apology and clarifying comment above.

but they're not actually equivalent, in spite of tables like this [0]. type ii error is a false negative result in the context of a test, where you have to understand which hypothesis is which and exactly what you are accepting or rejecting (hypotheses are not always as simple as hotdog/not-hotdog); if your listener doesn't know what statistical tests mean or wasn't following the setup, they have to stop you and ask.

[0] https://en.wikipedia.org/wiki/Type_I_and_type_II_errors#Tabl...

would you not expect your hypothesized (theoretical) ML expert to understand boosting, which is generally explained in terms of PAC learning, which draws on computational complexity?

that said, i'd also expect a phd in statistics to be able to figure out boosting without taking an undergrad course that worked up from automata. so the halting problem test, while it does capture something, may not be quite right.

ali rahimi alludes to the problem of google engineers simply needing to tweak models that were previously tuned by google researchers who do have well-developed intuition [0]. because the intuitions in explicit form are at best heuristic and not necessarily even consistent, signing up to improve a model without them might result in spending indefinite time and compute resources without guarantee of positive results. which is a terrible perf-theoretic strategy...

[0] http://www.argmin.net/2018/01/25/optics/

type ii error is statistics, not mathematics. there is no equivalent concept in CS because type ii error relates specifically to statistical inference and hypothesis testing.

that said, if you are just pointing to a box in a confusion matrix and saying "TYPE II ERROR," you are probably trying too hard.

i think the term is more prominent on the NLP side, via information retrieval (IR) and clustering. i first saw it in IR, and you'd see it in stanford's CS 124 or CS 224N, for example. if the parent is talking about people who are working on a system that has text-understanding component, i can understand their surprise.

i would request that you stop critiquing "machine learning" based on the presentation in introductory online materials like this and the ng coursera course. you provide a lot of signal in general but i think these critiques do decrease your SNR.

i am certain that you are familiar with the "usual" statistics sequence. (for others: there are lower-division courses that use calculus in a few places but otherwise avoid it, focusing instead on memorizing procedures. there is the upper-division probability/math stat sequence that uses calculus heavily but avoids analysis. and there is an intro phd sequence that finally gets into measure theory.) if you look at a coursera course that gives an high-level overview of practical statistics and simplifies its presentation to be accessible to people who never took calculus, you can criticize the very idea of a course that does not explain the measure-theoretic issues, but it makes no sense to use it to criticize the field of statistics, or to criticize the competence of others in the institution that produced it.

here, google is producing introductory training materials for developers. many developers have never taken calculus, let alone optimization, statistics, or analysis, and when i took MLCC internally, you were supposed to go through this whole thing (lectures and coding) in two days. it's supposed to give you enough understanding of the concepts to understand the API and apply it.

to the point mentioned somewhere else about outcomes vs. optics: some of the academic beauty pageant rankings beloved of university marketers do include nominal instructor qualifications (faculty status of instructors, fraction of faculty with phds, etc.) as metrics.

lack of security of employment makes teaching at low pay unattractive compared to industry at high pay. look at this posting [0]. nowhere does it say that this ladder is ever eligible for tenure, unlike (say) berkeley's lecturer SOE position.

"The Department of Computer Science at Stanford University invites applications for an educationally-focused faculty position. The responsibilities of the Assistant Professor of Computer Science Education will include: teaching classes (typically three or four courses, depending on class size, during the three quarters of the regular academic year), working to develop CS undergraduate curriculum, and involvement in the broader CS education community (e.g., research and/or leadership in CS education). The appointment has an initial term of three years with the possibility of reappointment of an additional three years. This position is eligible for promotion into the non-tenure (Teaching) faculty line. After completion of the five-year term as Associate Professor (Teaching), the candidate will be considered for promotion to Professor (Teaching). Experienced candidates with an already established reputation and national visibility in computer science education may apply for a senior appointment (Associate or Full Professor (Teaching)), commensurate with level of experience. Appointments at the level of Associate or Full Professor (Teaching) may be for a continuing term."

[0] https://cs.stanford.edu/jobs/faculty-opening

EDIT: interestingly, the "lecturer" position is at least eligible for SOE.

[1] https://cs.stanford.edu/jobs/lecturer-opening

yes - "...neither confirm nor deny..."

it was never a secret that many of the surface navy's combatants were nuclear-capable for the entire duration of the cold war - stuff like ASROC/W44, but also TLAM-N. as a DD SWO in the early 90s, i went through some of the same surety inspections as you guys.

(that said, i think that platform proliferation is a terrible idea, and since deploying such platforms anywhere but blue water makes little sense, you'd effectively be reducing the already marginal deployable strength of the current surface force. the SWOs have too much to cover already.)

ABET accreditation is for degree programs, not individuals. it sets curricular standards.

most credible engineering programs maintain ABET accreditation, as it has been required for graduates to become PEs [0].

[0] i recently heard that berkeley EECS is dropping its ABET accreditation, as this is no longer true: https://eecs.berkeley.edu/sites/default/files/abet_letter_to...

kudos.

ABET-accredited engineering programs have had to document instruction in design ethics for a long time. anybody designing a good CS curriculum should know that and be thinking about whether CS really ought to be different in that regard.

berkeley has offered CS 195 [0] for more than three decades.

[0] http://inst.eecs.berkeley.edu/~cs195/

i think you’re referring to “Zombieboard” (eric saund published some papers on it in the 90s). it was installed in most meeting rooms, not just a demo room.

if your interests tend toward machine learning, you may run into calculus again. in particular, if you keep going past the "train a neural network" phase (for which very basic calculus is fine) to the optimization problems that are under the hood of most ML algorithms, you wind up in the land of functional analysis.

it is (a standard HW problem). for bonus points, there's a connection between order statistics of the uniform distribution (such as max) and the beta distribution, of which the solution above is an example.

i know what you mean.

on the other hand, many of the problems in CS 229 are proofs and derivations that are very similar to those in the lecture notes (forcing you to understand the lecture notes in detail). as such, they can't just "tweak the problems every year" like you'd do in a lower-division math course - they're more like the exercises in an upper division math text (some problems are in fact referenced in the notes).

so, it is perhaps understandable why they might not want solutions floating around the internet forever.

(yes, i know that the stanford honor code is supposed to deal with this.)

the SEE materials are from 2007.

i suspect that what he's referring to is that he's heuristically minimizing a somewhat arbitrary (loss) function in a million-ish dimensions using the simple variants of gradient descent that work under these conditions. it sounds far too WIBNI to produce good results reliably (in practice, let alone in theory). the landscape has so many stationary points at which to get stuck; why would you ever get good results?

there's a small cottage industry of papers (like [0]) that try to explain this.

[0] https://arxiv.org/pdf/1412.0233.pdf

The Navy has been using such routing for decades.[0] Seems like the big win for Fleet Numerical would be in not having to run the workflow by hand [1] rather than adding (largely uninterpretable) neural networks to the problem...

[0] http://www.oceanweatherservices.com/featured_blog_posts/abou...

[1] http://www.navy.mil/submit/display.asp?story_id=27544,

am i the only one who thought: "hey, apple finally implemented the ANY key!" https://en.wikipedia.org/wiki/Any_key