berkeley eecs graduate admissions

This year (2014-15) I served as a student reader on the Berkeley EECS PhD admissions committee. Berkeley typically gets about N applications specifically for CS, of which around M are accepted, so reviewing the applications is a big job, and often agonizing in that it involves making impossible choices and rejecting some amazingly qualified people. There are about K1 faculty and K2 grad students on the admissions committee, so each reviewer sees about C applications.

The admissions process begins in earnest on December 22, the day when faculty recommendation letters are due (if a letter is submitted late, the application is not disqualified, but might be at a disadvantage if one or more reviewers read the application before the missing letter becomes available). Every application receives at least two reviews, one of which must be from a faculty member. Reviewers score applications on a five-point scale:

• 1 (2/100): rockstar, definite admit
• 2 (): very strong candidate
• 3 (): strong application, nothing wrong, but doesn't stand out
• 4 (): potentially successful in a PhD program, but not at berkeley
• 5 (): background not well suited for a PhD

Most admits are 1s and 2s (TODO: do 3's ever get admitted? what if raters disagree?). Again this varies by area: AI has a roughly 2% admissions rate, so if 2 of 100 candidates are 1s there is not a lot of room for non-rockstars. Of course there are other considerations, such as whether an applicant's research interests are a good match for Berkeley faculty.

Notes from the admissions meeting (12/5/14):

-- admissions are done by area (AI/THY/etc). Applications will be funneled to a different set of readers for each area, and admitted based on the allocation for that area. Some areas are much more competitive than others (e.g. AI), so your chances of getting in are better if you apply under another area. However, this is only viable if you actually intend to pursue that second area; it's not necessarily easy to switch from a less competitive area to a more competitive area. (see next).

-- Admissions requires a source of funding. At least 1 prof must guarantee a GSR. Sometimes only one prof will do so. So although we say, "you get admitted to the program, then choose an advisor", it may be that only one prof guaranteed to be willing to advise you, and other profs have promised their slots to other students. So your actual choice may be limited. This also interacts with choice of area.

-- Informal target of 1/3 international, 2/3 domestic, for financial reasons (international students require nonresident tuition). This is not a quota and can very quite a bit from year to year.

-- The department places a great deal of importance on recruiting a diverse class of PhD admits. The way this is framed to admissions readers is that we do not discriminate by ethnicity or gender, but we want the strongest possible admits, and in order to achieve this we need to consciously understand the ways in which applicants from disadvantaged backgrounds may be extremely strong while not matching a "typical" academic profile. For example, minority students may be more likely to have attended lower-ranked regional colleges with fewer research opportunities, not because they weren't good enough for MIT, but perhaps due to financial constraints, family obligations, or because they didn't have college-educated family members or prep-school guidance counselors pushing them towards elite institutions. Women may not have seen CS as an option until, say, taking an intro course in their sophomore year and loving it, and consequently might have had less time for upper-level courses. And recommendation letters should be read with consideration of potential biases, generally unconscious, on the part of the writer. With these considerations in mind, all applicants still have to meet the same high bar.

-- see the Powerpoint for more things to potentially comment on: how the reviewing process works (one grad and one faculty reader, scores from 1-5, comments, 150 apps read per grad student, more per prof, guideline ten applications per hour i.e. six minutes each), calibration of recommenders (faculty can see past recommendations), does GPA matter (do we keep any record of which schools are inflated?), how do we evaluate undergrad research (recommendation letter will often say what part they played), …

THOUGHTS ON REVIEWING:

• after glancing briefly at personal history (what school, degrees, work experience), first thing I check is rec letters. then transcript to make sure grades and background (CS/AI/math) is reasonable.
• SOP is mostly ignored at the first level of screening. A bad SOP can sink you (no mention of research, poor writing or obvious lack of English fluency), and a very strong SOP -- showing clear technical understanding of research ideas, describing past work and presenting a compelling view of future research interests -- can boost your case, but most SOPs are quite generic. They include some flowerly (and always ignored) rhetoric about how the applicant came to discover their passion for AI research, they give a basic overview of the applicant's academic background and research experience (essentially recapitulating the CV), and somewhere include the obligatory paragraph containing one sentence each about the severla Berkeley faculty they'd be interested in working with. This is totally fine if the rest of the application is strong -- not everyone has a compelling, specific proposal for a future research program (note the bar is a bit higher for students already in an MS program) -- but means the SOP pretty much gets ignored, at least at the first level of screening.

WHY PEOPLE GET REJECTED FROM BERKELEY:

• applying with an MS already, but without superstar research results during their MS program. The bar for MS applicants is higher, partly because they've had more time to show research potential, partly because consequences of failure are worse (if you fail out of a PhD program with a BS, we can at least give you an MS. If you already have an MS, you get nothing).

• applying to the MS program. we don't have an MS program, so these applications are reviewed in the PhD pile. Many of them get rejected as unprepared for a PhD, which is often exactly the reason the students wanted to do an MS in the first place.

• Weak bachelor school (applies to lower-ranked American schools as well as many international institutions). Many weaker universities don't offer courses in AI, ML, or advanced math. The students may be extremely intelligent, but without a strong course background they are not prepared for a top AI PhD program (they may still do well in a lower-pressure program that gives them some time to catch up on background).

• Hard to calibrate international applicants. Grading scales are different, hard to judge. Rec letters are hard to calibrate. The strongest American rec letters say things like "strongest student of 30 undergrads I've advised", "as strong as former students X, Y, Z who [went on to top programs / are now successful researchers]". Many foreign schools rarely send students to top American programs (something of a chicken-and-egg problem) so it's hard to make these comparisons. Many foreign applicants have done senior projects that, had they been supervised by a professor well-connected to the Western research community, might have been channeled into a paper published in an recognized international venue, but instead are unpublished or published in obscure regional venues of uncertain quality. At the first level of admissions screening, we generally don't have the time to read technical reports and judge research quality directly (recall that we budget six minutes per application), so these students usually get discarded as "insufficient evidence of research potential". (many of these also apply to applicants from lower-ranked American schools, though we get fewer applicants from those programs).

• Applying in an area that wasn't the focus of their undergrad research (e.g., applying in AI but with undergrad research in networking) without an extremely compelling story for why their interests changed and why they'd be successful in the new area (for example: did networking research sophomore summer, but took an AI class as a junior and loved it, now have taken a bunch of upper-level/graduate AI classes and am beginning an independent study in that area with strong rec letter from AI faculty).

• A good application that just doesn't stand out.

• The preceding are all reasons you might not make it on the shortlist. Of shortlisted candidates, quite a few qualified applicants will still be rejected due simply to a lack of research fit. If you're really into subarea X but we don't have any faculty who work in that area, no one will admit you. You're also out of luck if the professor you'd want to work with just isn't taking students this year, or is looking for a student to work on a particular project that you might not be a good fit for (note: even a professor who's "not taking students" can usually be persuaded if a truly exceptional candidate comes along, but the bar there is much much higher).

Conclusions: I would not have admitted myself to Berkeley. I probably wouldn't have let myself past the first level of screening.