How To Do Machine Learning Research Without A PhD
Summary
TLDRThe video script discusses the roles and requirements in AI research labs like Google Brain, DeepMind, and OpenAI. It clarifies that while a PhD in machine learning is often expected for research scientists aiming for top-tier publications, it's not mandatory for software engineers who build infrastructure. Research engineers fall in between, contributing to research papers and requiring less specialized ML knowledge. The speaker encourages those interested in AI research to consider applying for roles even without a PhD, emphasizing on-the-job learning and familiarity with ML jargon.
Takeaways
- š§āš» There are three common roles in research labs: software engineer, research scientist, and research engineer, each with different requirements for machine learning knowledge.
- š§ Software engineers primarily build infrastructure for research teams and typically require less machine learning knowledge, often not needing a PhD in ML.
- š¼ Software engineers might specialize in low-level systems or distributed systems, focusing on performance and scalability across data centers.
- š Research scientists are expected to publish in top-tier conferences and usually have a PhD, as their main output is research papers.
- š¬ Research engineers sit between software engineers and researchers, contributing to research papers and often implementing novel techniques from academic papers.
- š¤ Research engineers may work closely with software engineers and researchers on ambitious projects, utilizing more resources to tackle complex problems.
- š Not all research labs have a research engineer role; for example, Google does not, while DeepMind and OpenAI do.
- š Having a PhD is not a prerequisite for all roles, and even software engineers can publish papers in top-tier venues.
- š The speaker advises against pursuing a PhD solely for the purpose of working in a research lab, suggesting there are many other options available.
- š It's recommended to take introductory courses in machine learning to familiarize oneself with the field and the specific research focus of the lab of interest.
- š¼ For those interested in AI research, applying directly to labs as a software engineer or research engineer without a PhD is encouraged, as skills can be learned on the job.
- š The transcript emphasizes that a PhD is not always necessary for cutting-edge work in machine learning and AI research labs.
Q & A
What are the common roles found in research labs like Google Brain, DeepMind, OpenAI, and Facebook AI Research?
-The three common roles are software engineer, research scientist, and research engineer. Each role has different responsibilities and requirements for machine learning knowledge.
What is the primary responsibility of a software engineer in a research lab?
-A software engineer is mainly responsible for building infrastructure for the research team, which could be specific to a single research experiment or used by multiple teams for various research papers.
Do software engineers in research labs need to have a deep understanding of machine learning?
-Typically, the required knowledge of machine learning for software engineers is at the lower end, and they are less likely to solve complex machine learning problems during interviews.
Is a PhD in machine learning necessary for the role of a software engineer in a research lab?
-No, a PhD in machine learning is not typically required for software engineers in research labs.
What are the expectations for a research scientist in terms of academic contributions?
-A research scientist is expected to publish papers in top-tier conferences such as NeurIPS, ICML, ICLR, or other specialized venues related to their subfield.
How common is it for research scientists to have a PhD?
-About 99% of research scientists have a PhD in their chosen field or a closely related field to machine learning.
What is the role of a research engineer in a research lab?
-A research engineer sits between a software engineer and a researcher, helping with research papers, running experiments, and contributing to specific papers with novel techniques or ideas.
Do all companies have a research engineer role?
-No, some companies like Google do not have a research engineer role, while others like DeepMind or OpenAI do.
Can software engineers also act as research scientists and publish papers?
-Yes, many software engineers also engage in research activities and publish papers in top-tier venues.
What advice does the speaker give regarding pursuing a PhD solely to work in a research lab?
-The speaker advises against spending five years pursuing a PhD with the sole purpose of working in a research lab, as there are many other options available.
What should someone interested in working in a research lab do to prepare?
-The speaker recommends taking introductory courses in machine learning to familiarize oneself with the jargon and research topics of the lab they are interested in.
Outlines
š¬ Roles in AI Research Labs
The speaker discusses the varying roles within AI research labs such as Google Brain, DeepMind, OpenAI, and Facebook AI Research, emphasizing that a PhD in machine learning is not always required. They outline three common positions: the software engineer, who primarily builds infrastructure and may not need a PhD; the research scientist, who is expected to publish in top-tier conferences and typically holds a PhD; and the research engineer, who bridges the gap between the two, assisting with research and possibly contributing to papers. The video aims to clarify the misconception that a PhD is mandatory for all roles in AI research and encourages those interested to apply directly for positions such as software or research engineer roles even without a PhD.
š Preparing for AI Research Lab Roles
In this paragraph, the speaker advises viewers to take introductory machine learning courses to familiarize themselves with the field's terminology and to gain an overview of the type of research conducted by the labs they are interested in. They mention that while having a PhD is not a prerequisite for working in a research lab, understanding the basics of machine learning is beneficial. The speaker also cautions against pursuing a PhD solely for the purpose of working in a research lab, suggesting that there are other paths available. They conclude by inviting viewers to share their thoughts and questions in the comments section.
Mindmap
Keywords
š”Research Lab
š”PhD in Machine Learning
š”Software Engineer
š”Infrastructure
š”Distributed Systems
š”Research Scientist
š”Publishing Papers
š”Research Engineer
š”Cutting-Edge Research
š”Machine Learning Jargon
š”Reinforcement Learning
Highlights
Working in top research labs like Google Brain, DeepMind, OpenAI, and Facebook AI Research does not always require a PhD in machine learning.
Three common roles in research labs are software engineer, research scientist, and research engineer, each with varying ML knowledge requirements.
Software engineers primarily build infrastructure for research teams, which may be specific to one experiment or used across multiple teams.
Software engineers may specialize in low-level systems programming in C++ or work on distributed systems infrastructure for scalability across data centers.
Machine learning knowledge for software engineers is typically at the lower end, with less emphasis on ML problem-solving during interviews.
A PhD in machine learning is generally not required for software engineers in research labs.
Research scientists are expected to publish papers in top-tier conferences like NeurIPS, ICML, ICLR, and are typically required to have a PhD.
Research scientists collaborate with software engineers on ambitious projects and have access to more resources for tackling complex problems.
The research engineer role sits between the software engineer and researcher, assisting with experiments, infrastructure, and contributing to research papers.
Research engineers focus on novel techniques, implementing academic papers, and may propose research ideas for publication.
Not all research labs have research engineer roles; for example, Google does not, while DeepMind and OpenAI do.
Many software engineers act as research scientists, publishing papers in top-tier venues, indicating that a PhD is not a strict requirement for publishing research.
There is no need for a PhD to work in cutting-edge machine learning and AI research; consider applying as a software or research engineer without one.
Taking introductory machine learning courses is recommended to familiarize oneself with the jargon and research of the lab to which one is applying.
Pursuing a PhD solely to work in a research lab is not recommended, as there are many other options available for gaining experience and expertise.
The big takeaway is that a PhD is not always necessary for contributing to state-of-the-art research in machine learning and AI.
Transcripts
A lot of people have asked me what it takesĀ to work in a research lab like Google Brain,Ā Ā
DeepMind, OpenAI and Facebook AI Research. AndĀ a very common question I get is, do I need aĀ Ā
PhD in machine learning to work on cutting edgeĀ machine learning? I have news for you that isĀ Ā
maybe. So stay tuned and I'll tell youĀ which roles require a PhD and which onesĀ Ā
don't. So let's get into it. There's three veryĀ common roles that you find in research labs,Ā Ā
and they sort of exist on a spectrum of minimalĀ requirement for machine learning knowledgeĀ Ā
to requiring very deep specialized ML knowledge.
The first most common job role exists on thisĀ end of the spectrum, and that's going to be theĀ Ā
software engineer. This software engineerĀ is primarily concerned with building outĀ Ā
infrastructure for the rest of their researchĀ team, and this infrastructure might be specificĀ Ā
to a one research experiment for a single paperĀ or it might be some info that's used by multipleĀ Ā
teams or even ML info that's going to be used forĀ multiple research papers. These software engineersĀ Ā
typically need to be very strong at either lowĀ level systems if they're doing kind of low levelĀ Ā
programming and C++ or doing things onĀ like TPUs or GPUs at the kernel level.
Alternatively, you have software engineers whoĀ are more specialized in distributed systems, andĀ Ā
they're going to be building out infrastructureĀ that can scale across a cluster of machines thatĀ Ā
are available in data centers. Now, both of theseĀ software engineering roles are sort of specializedĀ Ā
in those particular domains and it's by no meansĀ easy work. But they're required knowledge ofĀ Ā
machine learning typically is going to beĀ at the lower end, and they're less likelyĀ Ā
during an interview to be required to solve theĀ kind of problems that other roles would require.
Oh, and one more thing about the softwareĀ engineer is they typically don't require a PhDĀ Ā
in machine learning, so that's one role inĀ a research lab that doesn't require a PhD.Ā Ā
Now on the flip side, on the opposite of theĀ spectrum, you have the research scientist,Ā Ā
the research scientist has an expectationĀ typically of publishing papers into topĀ Ā
tier conferences like NeurIPS, ICML, ICLR, or ifĀ they're specialized in some subfield like computerĀ Ā
vision or speech, they're going to want to publishĀ in those particular conferences and those venues.
Because they're expected to publish papers and doĀ research that is cutting edge, typically they'reĀ Ā
required to have a PhD. And I'd say about 99% ofĀ research scientists have a PhD in their chosenĀ Ā
field or very close field to machine learningĀ because their primary output is going to beĀ Ā
research papers. Many times you'll have a researchĀ scientist working with software engineers and theyĀ Ā
will be building out some larger projects thatĀ is more ambitious and requires a larger team,Ā Ā
but also they have the unfair advantage ofĀ having more resources at their disposal toĀ Ā
solving maybe a much harder problem thanĀ they could if they were just one person.
And then there's a third role, which is going toĀ be the research engineer. The research engineerĀ Ā
sort of sits in between the software engineerĀ and the researcher. Their expectation is to helpĀ Ā
with research papers. They're typically runningĀ experiments, helping out on the infrastructureĀ Ā
side, maybe applying it to a very specificĀ problem contributing for a very specific paper.Ā Ā
And they will typically be focused on veryĀ novel techniques or reading academic papersĀ Ā
that they downloaded off of archive, orĀ that they saw at our research conference.
And they'll take that information and try toĀ implement some papers. And they might evenĀ Ā
be proposing ideas for avenues of research inĀ their particular like in a paper that they'reĀ Ā
trying to publish with their research scientistĀ peers. Now, some research labs don't actuallyĀ Ā
have research engineers. For example, at Google,Ā there's no such thing as a research engineer role.Ā Ā
And whereas at companies like DeepMind or OpenAI,Ā Ā
that role exists and has a more clearlyĀ defined skillset. Not everyone who has a PhDĀ Ā
is a research scientist. There are many softwareĀ engineers who actually on their day to day jobĀ Ā
act as research scientists, and they areĀ publishing papers in top tier venues.
So you don't need to feel like because you'reĀ not a research scientist, you can't publish.Ā Ā
So yeah, there you have it. That's a high levelĀ overview of the different kinds of roles you haveĀ Ā
in state-of-the-art research labs, doingĀ machine learning and artificial intelligenceĀ Ā
research. The big takeaway is that you don'tĀ actually need a PhD to do a lot of this work.Ā Ā
In fact, if you're very interested in gettingĀ into that kind of work, strongly considerĀ Ā
applying directly to one of those labs as maybeĀ a software engineer or a research engineer ifĀ Ā
you don't have a PhD already, because you'll beĀ able to learn many of those skills on the job.
Now, I would highly recommend still taking someĀ introductory courses in machine learning so thatĀ Ā
you're familiar with the jargon or at least haveĀ an overview of the kind of research that the labĀ Ā
you're applying to does. For example, OpenAI doesĀ a lot of reinforcement learning type research,Ā Ā
and you should probably be familiar with thatĀ kind of a subject before you're applying. ButĀ Ā
don't feel like you need to have a PhD beforeĀ you do that. I would not recommend spending fiveĀ Ā
years of your life pursuing a PhD with the soleĀ purpose of wanting to work in a research lab.
There are many other options available to you,Ā and that can be the topic of the next video.Ā Ā
With that, I'd like to thank you for your timeĀ and let me know below in the comments anythingĀ Ā
5.0 / 5 (0 votes)