Part 02

What they worked on, and where it took them

Deep learning went from 1.3% of computing dissertations in 2012 to 50.1% in 2024, and most of that happened outside machine learning. What a graduate worked on then predicts where they land about as strongly as where they studied.

Doctoral dissertations 2005–2024 · subject and method read out of every abstract by a local open-weight model

This is the half of the question that career data alone can never answer: not just where computing doctorates went, but what they had spent four to six years working on before they went there.

Every abstract in the corpus was read by a local open-weight model and given a subject area and a set of method flags: does this dissertation use machine learning, deep learning, a large language model. The two charts below are doctoral dissertations judged core computer science, 58,998 of them from 2005 to 2024 for the subject chart and 48,436 from 2010 to 2024 for the method chart.

0% 4.6% 9.2% 13.8% 18.5% 23.1% 2005 2007 2009 2011 2013 2015 2017 2019 2021 2023 2024 Machine learning Computer vision Security & privacy Architecture & hardware Theory Networks
Share of doctoral dissertations by subfield, six largest areas.
0% 20.0% 40.0% 59.9% 79.9% 99.9% 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 Computer vision Machine learning Architecture & hardware Networks Theory
Share of dissertations within each subfield that use deep learning as a method. This is orthogonal to the chart above: it is not about which areas grew, but about a technique spreading into areas that are not themselves machine learning.

Machine learning as a subject grew steeply, from 5.9% of doctoral dissertations in 2015 to 20.6% in 2024. That is the visible half of the story and the smaller one. Deep learning as a method went from 1.3% of 2012 dissertations to 50.1% of 2024 ones, more than double the share that are about machine learning. So at least 59% of the dissertations using deep learning in 2024 sit outside the field: it became the default tool in areas that are not about learning at all. By 2024 it is in NLP (92%) and computer vision (89%), which is unsurprising, but also in 59% of computational biology and 43% of robotics, and still only 6% of theory. Dissertations using a large language model are the same curve seven years behind, at 17.0% in 2025 from nothing before 2018.

Theory 29.1% HCI 36.3% Security & privacy 28.3% Networks 22% Machine learning 23.4% Computer vision 17.1% Computational biology 27.6% IR & data mining 22.2% Architecture & hardware 12.4% Languages & SE 20.5% Systems & distributed 16.2% Robotics 20.6% NLP 18.6% Scientific computing 20.9%
Share going into academia one year out, by dissertation subfield, 2015–2019 cohorts.

Among 2015–2019 doctorates, architecture and hardware sent 65% into industry and 12% into academia. HCI is close to the mirror image at 41% and 36%. Measured on a single axis so the comparison is like for like, subject spreads destinations 23.9 points across 15 subfields, against 23.4 points across 54 institutions. What you worked on matters about as much as where you studied.

The ordering is not prestige. It tracks how directly a body of work maps onto something a company already ships.

Which subjects hold on to their academics

A year after the degree, a postdoctoral position and a faculty position look the same. Following the 2015–2019 cohorts out to year five separates them.

Dissertation subject Doctorates In academia, year 1 At year 5 Change, points
HCI 479 36% 33% -3.3
Theory 645 29% 22% -6.9
Security & privacy 568 28% 25% -3.3
Computational biology 500 28% 23% -4.8
Machine learning 668 23% 18% -5.9
IR & data mining 473 22% 17% -4.9
Networks 726 22% 18% -4.0
Scientific computing 258 21% 16% -5.0
Robotics 393 21% 16% -5.1
Languages & SE 425 20% 17% -3.6
NLP 328 19% 16% -2.7
Computer vision 884 17% 14% -3.2
Systems & distributed 520 16% 14% -2.7
Graphics & visualisation 244 14% 12% -1.2
Databases 137 13% 14% +0.8
Architecture & hardware 756 12% 12% -0.5

The range is the story. HCI keeps 36% of its doctorates in academia at year one and 33% at year five; architecture and hardware is at 12% and 12%. That is a factor of 2.9 in how likely an academic career is, decided years before anyone applies for a job. The two ends are fixed: across all three cohort eras in the data, HCI is first and architecture and hardware last every time.

And the movement runs one way. Fifteen of the sixteen subjects have a smaller academic share at year five than at year one, the sole exception being databases at +0.8 points on 137 people. Nobody comes back: the academic share a subject has at year one is about the most it will ever have.

Two things a student can act on. Deep learning is no longer a specialisation, it is the default method in half of all computing dissertations and in most of the applied ones. And the subject chosen at the start of a PhD sets the odds of an academic career by a factor of nearly three, before anyone applies for anything.