A Research Blueprint

Finding your Research

Shruthi Kubatur BMS College of Engineering April 2024

Sr. ML Research Scientist, Apple Inc.

PhD (EE), Purdue University

 Something
the world What’s
your
needs

calling?

Your ideal Something

Something research you can get
you love topic paid (well)
for

Something
you are
good at
But… Something the
world needs — but

Practical
what else is left
after large
language models?

Challenges

Something Your ideal

you love — research topic Something you
may not — may not can get paid
have even exist! (well) for —
prospects perhaps after you
graduate

Something you’re good at

But researchers are happy, right? There’s a
— may be just a small trick to it…
speck of what’s needed
 • Read articles — they hand a lot of important
Something problems to you on a silver platter
the world
• Read papers — this needs more work
needs sometimes, but read papers like you are
reviewing them for faults
• Some of these shortcomings can be genuinely
fascinating
This is where it’s
important to read
and play • Some of these can t right into your dissertation topic

• Play with technology, gadgets, algorithms

• Practical experience > just theoretical knowledge
• Try to break them — maybe you can x them in your PhD
• Try to apply them — maybe you can use them for
something the creators had not originally intended for
fi
fi
 Choosing Papers to Read Pick papers with
at least K citations

Choose a topic that’s < 1 yr 1-4 yr 4+ yr

neither too broad nor too old old old

speci c Regular
paper
Speci c enough to keep you focused (good 1 2 4
Broad enough to give you space venue)
Survery/
Choose something that’s review 3 5 5
paper
neither too old nor too new
arXiv
New enough to be relevant paper 2 4 5
Old enough to be mature and cited
Stellar
author / 0.5 1 2
venue
Pick at least 4 di erent
papers per topic Multiplier: mean citations in
that topic (mean computed
Diversity is key over mature papers)
But too many papers can be confusing
Note: some elds are faster
than others (e.g. AI/ML is
Pick papers from a good faster than pure math). So
calibrate the time periods
venue (IEEE Trans., accordingly. The above table
NeurIPS, CVPR, some shows an example for AI/ML
Springer journals). Beware papers.
of low-standard venues.

Disclaimer: This is a general strategy that can be used while picking research problems; feel free deviate from this for your speci c situation.
fi
fi
fi
ff
fi
 • Life shouldn’t be a struggle — looking at
Something undergrad classmate’s mega mansion
shouldn’t hurt deep inside
you can get
paid (well) • Correlated with utility — ability to get a
good-paying job means working in an
for area that is probably useful
• Not always true, but mostly is
Doesn’t mean
you’ll end up • In capitalistic view, correlates with what the
selling your soul world needs (De nitely not as the primary driver)
for stock options;

Or that you
• Ability to attract funding
should quit
physics to be a • For faculty aspirants: you need to feed grad
Wall Street trader. students, even if they can’t still buy organic milk

• For research aspirants: you still need funding

• For “start-up stargazers”: VCs are picky
fi
 • Be objective — evaluate your skill level without bias
Something
• A trick: imagine you’re interviewing a candidate with your exact
you are good skillset and knowledge
at
• Work in progress — ability is developed
• Be reasonable with where you want to get (mastering Python
may take a month; mastering signal processing may take years)
It may not be • But get to 80% fast
enough; but it’s a
good starting
point • Makes research more enjoyable
• It’s easier to push the boundary if you’re to the boundary
• Helps immensely with identifying interesting problems to work
on

• You’ll be more creative with your solutions

 • Don’t settle
• Sometimes PIs don’t have money for what you love
Something • Look for programs where there are multiple PIs working on similar topics (what exact topic they’ll
you love have funding for is a bit stochastic and time-varying)

• Try to get your own funding (grants, TA)

• Maybe work with a PI who has money for something you actually love (see next point)

• Really, what is it?

This is where it’s • Find out what you truly love about it
important not to • If you think you love cryptography, you may really be in love with number theory. So maybe do
compromise too math (?)
much
• Correlated with ability
• Caution: love and infatuation are two di erent things (like with people) — three straight lecture of
quantum loop gravity from MIT doesn’t mean it’s your love

• If you truly love a eld, you’re typically good at it

• Maintaining motivation and focus

• It’s important to remind yourself why you’re doing your PhD — that’s much easier if you love the
subject
fi
ff
 Something

If you’re curious, the world

needs

here’s my blueprint Better

healthcare

Better
material
design

Machine
Math
learning
Rare-event Something I
Something Signal
modeling + Algorithm could get
I love processing simulation design paid (well)
for
Coding
Coding
Math

Physics

Signal
processing

Something
shruthi@alumni.purdue.edu I am good
at