October 3, 2021 • 46 mins
In this episode, we sit down with Mihaela Roșca, a Research Engineer at DeepMind, one of the world’s most advanced and influential artificial intelligence research labs. DeepMind is renowned for its bold, high-impact breakthroughs—from mastering complex strategy games like Go and Starcraft, to revolutionizing protein folding with AlphaFold, and optimizing weather prediction and energy efficiency at planetary scale. It’s not just about research for its own sake—DeepMind’s work is actively shaping the future of science, technology, and society. Mihaela’s journey to DeepMind is a testament to the rigorous academic and applied excellence the organization represents. She began her career as a Software Engineer at Google Zurich, working on natural language processing using neural networks. She went on to pursue a PhD at University College London (UCL), under the mentorship of renowned machine learning professor Marc Deisenroth. Her research sits at the intersection of generative models, reinforcement learning, NLP, and the challenge of building scalable and safe machine learning systems—with a special intellectual curiosity for probability distributions and their central role in learning. Mihaela holds a first-class honours MEng in Computing from Imperial College London, where she was recognized for her academic excellence, and today she is part of a world-class team pushing the theoretical and practical limits of artificial intelligence. In this episode, we explore:
Back in 2021, while much of the industry was still catching up with the possibilities of AI, ReMotive and LA PIPA IS LA PIPA—under the leadership of Alex Lawton—were already operating at the highest standard of global thought leadership. This conversation with Mihaela Roșca of DeepMind is further proof: bold, technical, visionary, and years ahead of the curve.
- Her path from Google to academia, and from there to DeepMind.
- The role of foundational research in unlocking real-world applications of AI.
- Why probabilistic reasoning remains a critical area of focus in machine learning.
- What it means to build AI systems that are both powerful and safe.
- How DeepMind thinks about the societal implications of its work.
Back in 2021, while much of the industry was still catching up with the possibilities of AI, ReMotive and LA PIPA IS LA PIPA—under the leadership of Alex Lawton—were already operating at the highest standard of global thought leadership. This conversation with Mihaela Roșca of DeepMind is further proof: bold, technical, visionary, and years ahead of the curve.
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Speaker 1 (00:06):
Bedroom is an independent data science and aim specialized in
data driven business change. In this podcast, our guests help
us spread knowledge and experience with our listeners.
Speaker 2 (00:29):
Good morning, Mihaila, how are you doing today?
Speaker 3 (00:32):
Good morning, I'm great. How are you?
Speaker 2 (00:35):
I'm fine. I believe that the autumn is officially here.
It's getting older than ever here in the earth of Spain.
How is it in London?
Speaker 3 (00:45):
It's good.
Speaker 4 (00:45):
It's actually I was looking outside before the recording it's
quite sunny. But I also feel like, yeah, the summer
has ended. I think Saturday we had our last real
summer day.
Speaker 3 (00:56):
It was really warm and sunny.
Speaker 4 (00:58):
But I think it's yeah, we're we're getting braced for
autumn and winter.
Speaker 2 (01:04):
Are you a summer person or more like you like
winter and the Christmas lights? What do you prefer?
Speaker 3 (01:10):
I am a huge summer person. I love summer.
Speaker 2 (01:13):
Me too.
Speaker 4 (01:15):
I yeah, I really don't like winter, So for me,
this is a bit sad summer is over.
Speaker 2 (01:24):
Same here, okay, So tell me what do you have
in the agenda in the schedule for today? Is it
packed with meetings?
Speaker 3 (01:34):
Nos?
Speaker 4 (01:34):
I mean I always start my day with a green tea,
so I have it here with me and after we
talk today, I'll have some time for a little bit
of what I like to call focus time, so I'll
be thinking a bit about my next steps research wise,
and then i'll yeah, I'll have one or two meetings,
but yeah, not too much today. I think in general,
(01:56):
it does vary a lot from time to time where
they are. I'm sorry, I just spent my whole day
focused on paperwriting or doing some proofs. But now it's
a little bit of that period of trying to yeah,
reassemble and figure out what to do next.
Speaker 2 (02:13):
So whoever is listening to this here or see, we'll
hear about papers like you just mentioned. So I think
it's appropriate to provide some context on who you are,
where you come from. So youar, Mihaila, you're currently in
Deep Mind. You are working as a research I wouldn't
say engineer, right, how would you describe it? What's the
(02:35):
what's the what's the role?
Speaker 3 (02:36):
Yeah?
Speaker 4 (02:36):
My role is as a research engineer, that's right. So
that's I'm a stuff research engineer and I've been working
at deep Mine for yeah, five years in this strole
and at the same time, I'm also doing a PhD
at UCL.
Speaker 2 (02:51):
Thank you, And we have you here because we did
a podcast recording also with Cercio, with Sergio as you
may call him, who who works in deep Mine in
London and who was also part of Data Stand Up
Spanish podcast earlier this year. Right, he was the one
(03:11):
who told you to take part in this.
Speaker 3 (03:13):
That's right, that's right.
Speaker 4 (03:14):
He's an amazing colleague and a really nice person as also.
I was really happy when he reached out and asked
me what I want to do this and.
Speaker 2 (03:25):
We're happy to have you here. Okay, so you completed
your bachelor and your master's in London. I guess something
that's interesting to anything, to anyone listening to this conversation
would be what did prepare you well for your current role?
Because you're in a deep tech and a research firm
(03:46):
like deep Mind, And what did you miss maybe theoretical
knowledge or any sort of a practical experience before joining.
But yeah, what did prepare you well? And what didn't
sort of say?
Speaker 4 (04:01):
Yeah, this is a very good question. I think my
degree really focused on providing us with the good technical
expertise and also good knowledge on software engineering, but also
we had a lot of advanced courses on mathematics and
machine learning.
Speaker 3 (04:14):
Which had been extremely useful.
Speaker 4 (04:16):
So I think on the technical side, I was very
well equipped after finishing my degree. But I think one
aspect of research that is really hard to learn unless.
Speaker 3 (04:26):
You're really doing it.
Speaker 4 (04:28):
So it's kind of hard to teach us prioritizing a
lot of the research ideas that you or the team
might have. I think maybe earlier in life, I thought
all research ideas are really hard to get, and then
once you have one, you just spend your whole time
pursuing that. But in practice, we often have many research
ideas and a lot of the hard part of my
(04:50):
job is to also try to assess, okay, which are
the right research ideas to chase, and that depends on
risk level for that idea, but also or how promising
the impact of that idea is. So this is something
that you always kind of constantly have to balance and
it's a bit hard to teach. And the skill that
(05:10):
I found that I had to really learn on the
job and become better and better at it in time, and.
Speaker 2 (05:17):
So that balance is somehow support it. I mean that
decision that has to be made on what you focus
on is somehow supported by alphabet because for the listeners
that do not know, which I guess would be only
a few. But the mind is the AI and research
unit of WILL, which ultimately belongs to Alphabet. But that's
(05:41):
the business as we will not imposed but guide you
through the different fields that would be of interest for
you to pursue or do our research on something that's
your interest. How does it work, Well, we do.
Speaker 4 (05:55):
Have a lot of research freedom, so I think it
might operate relatively independently. But of course it's a matter
I find of trying to also convince your colleagues that
if you have an idea, that's worth pursuing, because often
you might have an idea. But again coming back to
(06:16):
what I've previously said, there are many ideas. So I
don't really work and avoid I work with my colleagues,
and it's a matter of together between all of us
trying to figure out what's the right thing to focus
on right now.
Speaker 2 (06:30):
So you're completing your PhD at the moment, right what's
the field of a study?
Speaker 4 (06:34):
Yeah, right now I'm in my PhD. I'm generally focusing
on learning probability distributions, and I'm really interested in learning
probability distributions and learning them efficiently. Because this is really
at the core of machine learning. So classification, generative modeling,
reinforcement learning, they all rely on learning distributions efficiently. In classification,
(06:56):
well learn a conditional distribution label given the data. In
generative modeling we might learn an unconditional or also conditional distribution,
and in URL in reinforcement learning, you learn a policy
which is also distribution condition on the state of what
action you should do next. So this is why I
find learning probability distribution is extremely important because if you
(07:20):
make progress here, you can make progress in a lot
of the fields. And I'm focusing a lot on what
kind of loss functions one should use to learn these distributions,
but also most recently a also how to optimize and
regularize neural networks that are learned to represent these probability distributions.
Speaker 2 (07:40):
I'm curious, then, so you've mentioned lost functions. I guess
that your work, as it's part of a research exercise,
it goes beyond yes, coding and creating a library that
summarizes this for prefers, this for any other inducer. But
it's about doing that mathematical research, not on paper, but
(08:02):
you have to do that deep work on what you
change in the back end sort of say right.
Speaker 3 (08:08):
Absolutely, absolutely. Yeah.
Speaker 4 (08:09):
So for example, in one of our recent papers, we
use mathematics, as you've just said, to try to understand
what gradient descent actually does in two player games, and
we can quantify that. For example, in our case, we
use an ode or ordinary differential equation, and based on
the math, we then see, oh, we can treat this
(08:30):
algorithm that then improves the performance.
Speaker 3 (08:33):
But yeah, there's a lot of new math.
Speaker 4 (08:36):
Let's say that describes a particular situation. In this case,
it was great and descent with games. And then we
take that math and we do something with it. And
the something often involves coding an algorithm so that we
can test our mathematical predictions empirically.
Speaker 2 (08:54):
Okay, we can talk about this later because I'd be
curious to know if these pre col demonstrations also go
to the business side of Google, for instance. Okay, back
to a few years back. I'm curious, what was the
Mind working when you joined back in twenty sixteen.
Speaker 4 (09:12):
Yeah, I mean the Mind was working on a lot
of really interesting things and a lot of the core
methods we still use today.
Speaker 3 (09:18):
So there was a lot of work.
Speaker 4 (09:19):
On reinforcement learning, generative models and also enhancing, for example,
reinforcement learning methods with memory retrieval and neural systems. And
I think, what's what's interesting when I saw this question
is looking back, how a lot of these core methods
are still really used today in our models.
Speaker 2 (09:39):
And so he told me, you're part of a research
team now trying to look at the big picture. And
probably you're familiar with the structure of the organization. How
many units is the mind composed of? Somehow what is
each of these ones doing?
Speaker 4 (09:59):
Yeah, so I am part of the research team, and
deep Mind also has an apply team, which focuses a
lot on bringing deep Mind research a lot of the
research that I talked about before to the world into products,
for example, and.
Speaker 2 (10:15):
If there was a need to implement those algorithms that
you've developed and you've tested into a product that I
don't know is part of the Google Cloud platform suite,
that be this apply team, right.
Speaker 3 (10:29):
Yeah, it depends.
Speaker 4 (10:31):
It can also be shared, so for example, the research team,
if they come up with an idea, they can also
collaborate of course with the apply team.
Speaker 3 (10:39):
Everything.
Speaker 4 (10:41):
One thing that I really like about depind is that
everything is flexible, so you can be part of the
research team, but if you have come up with an
idea and you really want to see that into a product,
you can still work on that.
Speaker 2 (10:53):
And when you joined, how many were you? I'm just
asking because I started to know about deep Mind when
the famous documentary of Alpha Go appeared in my Netflix
recommendation list and I watched it. I loved it, actually,
so I'm curious to know how did the company grow
(11:14):
from when you joined till today? How did you leave
this growth?
Speaker 4 (11:21):
Yeah, I mean I think I don't remember exactly how
many we were, but certainly there has been a lot
of growth scenes and I think the what that allows
the company to do, and I think what you can
see is focusing on more research directions and more very
(11:42):
ambitious products and projects. So, for example, as you've seen recently,
there was the Alpha fold result where a deep Mind
was able to produce a machine learning method which is
very good at protein folding and really advancing the state
of the art and science. So I think this this
is primarily what growth has achieved. I think the culture
(12:05):
and the atmosphere is very similar to the one that
I experienced when I joined in two thousand.
Speaker 2 (12:11):
Sixty, which I guess is one of the hardest things
to keep right. I mean, when you grow that much,
you add different people to the team, making sure that
those values, those initial values stay, that's really important as well.
Okay for anyone listening to this conversation that maybe feels
ambitious about joining deep Mind because he feels or she
(12:34):
feels that he wants to be a part of that
research mission that you're mentioning because that's her interest. What
do you look for in candidates? Have you ever take
part in hiring selection processing, and even if you didn't,
what do you think are the best traits to be
a good contributor to the Mind's work.
Speaker 4 (12:53):
Yeah, so I have been part of the hiring process
at various stages. I think the exact requirements will depend
on the role. So there are different technical interviews for
software engineering role, for a research engineer or for a
research scientist. But I think in general we are looking
for candidates who are just interested in artificial intelligence and
really like solving different problems because that's a lot of
(13:16):
the day to day of what we do, and also
are keen of being part of the defined team.
Speaker 2 (13:23):
And I guess that one of the most important things
when you are part of this cutting edge field is
to be up to date, read a lot, make sure
that you pay attention to everything that it's being published.
Do you keep time in your agenda in your schedule
(13:43):
to go through some of these papers. How much time
of your time relates to reading and consuming information from others.
Speaker 4 (13:51):
Yeah, I do spend quite a lot of time reading
and consuming papers, and I think it's really important to
stay connected with the research ideas from the community. I
also really like to take a few days just once
in a while to just read, so basically to block
out my calendars or to speak and just say, okay,
I'm going to now immerse myself into what's going on
(14:13):
either and through something that I'm researching right now or
something that is maybe somewhat related and I'm trying to
learn about. And I think, yeah, it's quite important for me.
And I think a lot of ideas and inspiration comes
from us. I mean, what's going on in the literature.
Speaker 2 (14:29):
I completely agree. I mean some of the summer when
I was taking some time off, I was going through
some of the articles that you can find in platforms
such as medium. When you select machine learning, AI or
data science, you can find some of the I mean,
those aren't as deep a paper, but you can find
(14:51):
some applications that provide you with ideas. And I believe
that taking some time off to reflect on what you're
doing us that we do. I mean, part of our
work is research. Even if we are a putique practitioner
of data science and AI for our clients, we still
need to stay ahead of what the competition is doing
(15:14):
and to make sure that we provide what's the latest
advice advancements in technology and applications. So I guess that
reading and consuming information is critical for professionals, regardless of
you know, being research or not. Okay, so you work
in a research field that, as I said before, is
(15:35):
categorized as categorized as cutting edge, but you get to
see these developments applied in real life sort of say,
because you've mentioned alpha fold, but did you ever encounter
a technology that you use in your day to day
that you think, Okay, part of the work that we
(15:56):
did in deep Mind is somehow embedded into the device
or this application, and how does it feel? Yeah?
Speaker 4 (16:04):
Absolutely, I think there are many examples of cutting edge
research which are now part of alphabet technology. So for example,
wave Net, which was a research project, is now part
of a lot of the text to speech products at Google.
And another nice recent example is traffic prediction as part
of Google Maps using graph nets. So there are many
(16:28):
others as well. So I think this is also what
makes machine learning research so appealing, is that you can
have these research ideas and actually they do lend themselves
to being very useful in practice. Then they can be
used by millions of millions of people.
Speaker 2 (16:46):
So when I'm using Google Maps, I should think of
the mind developing those predictions of what's the traffic going
to be like in my city? Correct?
Speaker 4 (16:58):
Yeah, yeah, it's part of the prediction and the recent
traffic Yeah, amazing.
Speaker 2 (17:05):
The thing is how how does the algorithm work in
that sense, because I guess that the system is being
fed continuously with the traffic information from all of the
different cities around the world. Is that information processed within
your device or how does it happen?
Speaker 4 (17:20):
So I'm not aware of how this actually works on
the on the application level. Okay, So yeah, I don't
know exactly how does is implemented in the application.
Speaker 2 (17:31):
Okay, So you've mentioned alpha fault before, and I also
mentioned that the mind somehow I started to become very
widespread and everyone started to knew about Alpha Go when
it was released in Netflix because of your developments for
(17:53):
the Alpha Go game, for the Go game. Sorry, so,
can you tell me something about this that we may
not know in terms of I don't know developments. I
wouldn't say secrets, but probably I don't know anecdotes about
the development of Alpha Go software or the Alpha Hool
(18:14):
development that could be of interest to the listeners.
Speaker 4 (18:18):
Well, yeah, I don't know if there's anything. I think
one thing that is really nice about these projects is
seeing kind of the excitement around the office, and I
think that's very hard to replicate or to understand. I
think I joined in Mind twenty sixteen, so the lisad
all Alpha Go matches had already happened, but there were
some further matches in twenty seventeen, and there was so
(18:41):
much excitement in the office around these matches, and I
think that really, yeah, it really has an impact on
your day to day to kind of come to work
and to see all of that unfolding life. And I
think it's also similarly the case with Alpha fail to
know that this amazing groundbreaking research is happening, and to
(19:03):
have chats with people about it and all of that
really keeps you very, very excited.
Speaker 2 (19:08):
Something that amazed me in regards to that when I
watched that documentary is that your colleagues were excited but
at the same time a little bit afraid that the
software that they developed, those algorithms, we're going to surpass
(19:28):
the human capabilities of famous players and very capable players,
And they were excited about their development being successful, but
at the same time they were feeling bad about the
human on the other side. I'm not sure if if
you saw that.
Speaker 4 (19:46):
Yeah, I mean, it's always interesting, right because we are
all humans. It's yeah, it's a bit of it is
an interesting human experience to see that you you're able
to build algorthon, which is very good, which is of
course promising because we're hoping to apply all of these
(20:11):
machine learning advances to make the world a better place,
which is obviously quite quite exciting. But also on the
other hand, as you said, it's hard not to have
empathy for the human who is the best, let's say,
go player in the world. So yeah, it's a bit
it's an emotional mix, but I think still nonetheless quite
(20:33):
quite an exciting experience to have.
Speaker 2 (20:36):
Yes, And related to that, well, something that we do
ourselves studies somehow related to that is when we start
to work with an organization where we have to reinforce
or promote data culture, where we are developing AI mechanisms
for the humans at that company to be used, is
(20:57):
that these mechanisms hurt created, designed and implemented to empower
the people at that organization and to not substitute them,
because that's sometimes a mix of feelings that people have,
right like, they see AI as something that could be
(21:17):
evil and substitute in some of the capabilities that these
guys have or you know that of the things they
are doing in their day to day, But it's more
about empowering those individuals to make the world a better place,
if you want to see it that way. That's how
we think and that's why we try to promote. But
(21:38):
at the same time, yes, of course, when you create
something that has strong capabilities and probably can make decisions
faster than a human, you have this feeling that is
overwhelming sure to say. Okay, so tell me about some
of your achievements in research or even in general, how
(22:00):
they impact it your life. On one side, and maybe
your work in alphabet on the other, in the mind
as part of alf IT on the other.
Speaker 4 (22:12):
Yeah, I mean what really drives my research is this
motivation that I personally have to understand what's going on.
So a lot of what I've done is this kind
of research that focuses understanding the mechanism by which certain
methods work or don't work. So, for example, we might
know that there are certain limitations of sub types of methods,
but we don't really understand why it fails in that way.
(22:34):
So I'm really interested in understanding that, and I've done
that for a ganzo, a type of generative model that
learns to generate, for example, images using a two player games,
but also variational inference and stochastic gradient estimation, and more recently,
as I mentioned, optimization in games and regularization in neural networks.
So that's the understanding part. And then once you have
(22:57):
this understanding part, you can really use that. Either myself
or other colleagues have done that taken that understanding to
improve these methods in parts of either gans or variational
inprints and so on. So this is on the research
side and on the coding side. I also been part
of designing research libraries that have been used and defined
or more widely at alphabet.
Speaker 2 (23:17):
Okay, when you talk about regularization in games, does it
relate to I don't know about or a decision that
has to be made but was never made before is
somehow correctly done. Can you tell me more about this,
explain it a little bit more.
Speaker 4 (23:35):
Yeah, So here I met to regularization the context of
neural networks. So often when we played these games, we
still implement our two players these games as neural networks.
So in the case of gans, for example, you have
a discriminator which is going to be a neural network,
and a generator which is going to be a neural network.
And often when you train these models, we see that
there are certain methods that basically regularize the neural network space.
(24:00):
The functional space that represents the discriminator or de gerator
can be very useful at improving performance. But sometimes we
know why and sometimes we don't. So for example, when
I say regularization, I mean sometimes let's say L two
regularization when you add this term to the lost function too,
(24:21):
let's say, make sure to penalize the L to norm
of the way. That's a type of regularization, But these
days we might have more complicated approaches or more complex
approaches to regularize these models. And for example, one that
I was particularly interested in is smoothness with respect to inputs,
So that says, I want on your own network that
(24:42):
if I change my input a little bit, my output
doesn't change too much, which is kind of an intuitive
thing to want from your model.
Speaker 2 (24:51):
That's really interesting. Got you okay? And living aside all
that technical research works sort of say, I think last
year and the previous one was a bit of a
weird here in regards to the pandemic and so on
and so forth. How did this impact your work? What
did you learn? How are you utilizing these learnings?
Speaker 4 (25:14):
Yeah, this is a very good question. I mean, of
course it affected everyone. We went proworking in the office
and seeing everyone every day to working from home. And
also I think what the pandemic has particularly done for
me is reminded me how important personal circumstances are, even
(25:34):
at work and for work, and how we really have
to be mindful that our colleagues have different personal circumstances
and challenges. And I think this was really made obvious
by the pandemic. Some people were alone at home, which
was obviously very hard. Some people were struggling because they
had to work and take care of their children, and
(25:55):
there are many made there's the variety of experience of
how this was felt was quite large, and I think
it really made us more mindful that even though maybe
one person is doing well, another person is doing not
so well for different reasons. And I think we really
have to account for that in our day to day
(26:15):
work and support each other. And I'm planning to stay
a lot more mindful about this as the COVID situation
hopefully gradually improves.
Speaker 2 (26:26):
So it's about feeling and showing that empathy to co worker,
colleagues and even friends. Right, Like, you don't know what
the other person may be going through or feeling at
a specific point in time, so you just need to
be with her with him, right.
Speaker 4 (26:42):
Absolutely, absolutely, And I think sometimes what I've also noticed
is that because of the pandemic, maybe people are more
likely to share that something is going on or the
situation is hard, which makes it easier for example, to
be there for them and to help. And then I'm
I'm hoping that we will continue having that that culture
(27:05):
of sharing and saying actually things are not so good
right now, so maybe we can.
Speaker 3 (27:10):
Yeah, we can adjust our workload or.
Speaker 4 (27:13):
Anything that we need. Sometimes just talking to someone can
can really help as well.
Speaker 2 (27:19):
I'm curious. I didn't think of this till now, but
did you, as you know, deep Mind, work on anything
in regards to fighting off the pandemic or its effect
for COVID? Was it any Was there any development in
regards to that.
Speaker 4 (27:37):
I don't really know myself, so I I haven't been
involved in in this directly.
Speaker 2 (27:46):
Yeah, I didn't know either. I mean, this was just
out of the blue. As a curiosity. We did something
ourselves that wasn't deep dech, but it was hard because
we well, i'll explain to you what we did. We
created uh recommendation engine for people that wanted to travel.
So depending on the location where they wanted to travel,
(28:09):
we would be retrieving and gathering data from different sources,
some in regards to COVID, some in regards to other
security aspects of that location, and providing the end user
or the person with a recommendation index about how safe
it was to travel to that location. It wasn't only
related to COVID that I think it was more related
(28:31):
to rebuilding and regaining confidence of the people you know
that needed to travel or that had to travel. And
also because I think or we thought this could have
an impact on the economy, we called this this tool
check check doto. And this is not me doing any
(28:51):
promotion here of the tool, because this was something that
we did provano sort of say, just because we felt
it could have an impact after the after COVID nineteen.
I completely assume that probably I don't know, the Mind
did some sort of COVID prediction in regards to how
it was going to be I don't know, moving around
(29:16):
London or the UK, I don't know. Okay. Now, living
aside your work at deep Mind and yourself even do
you also follow on pay attention to your biggest allies
or the say, in terms of making progress in the
field of AI, which probably is open AI and their
(29:39):
advancements such as CBD three and other developments. Because you've
mentioned that you were talking about text generation before, I
think and somehow it's related to this, so probably this
is of your interest as well, right DBD three.
Speaker 4 (29:55):
Yeah, I mean I do follow advances in the field,
generally both from academia and other result slabs and so
of course I have read and been extremely impressed by
the results of GIPD three.
Speaker 2 (30:06):
Do you do you expect that at some point as
as humans will be able to communicate completely openly and
fluently with computers, because I know that the hardest thing
here for computer is to improvise right, generalize and be
able to make decisions on the goal in the in
the in the worlds to be chosen and the words
(30:28):
to be used. And how far do you think we
are from this? Because one thing is you know, text
generation and chatbots. But at some point, I guess we
expect a complete fluent conversation with a human, even discussing
even feelings. You think we will get to some point
We will get at some point of this.
Speaker 4 (30:48):
Yeah, I mean I am one of the people that
actually believes that we will reach general AI, and I
think if we think every reach General II, then we
will also have a good converse way to converse with
these agents that are acting in the world alongside us.
Speaker 2 (31:10):
Okay, so you're a film believer that we will reach
general AI. So humans that can feel emotions, that are
able to generalize in problem solving, in movements, in everything,
do you think this will exist at some point, So.
Speaker 4 (31:24):
It's not humans agents. I mean I think I think so. Yeah,
I think we will. I'm not, let's say, one hundred
percent sure, but I think I'm generally thinking towards us
quite confidently, because I think so. Our brains are incredible
computation machines, and they have been able to produce high
(31:48):
levels of intelligence. And while our computational models that we
use right now in our computers are different than those
in the brain, the really nice thing is that we
actually get some really useful hints from the brain of
what we need to do to build these general agents
which can solve multiple tasks. So when I mean generali,
I don't necessarily think about feelings necessarily, but an agent
(32:11):
that like us can solve a variety of tasks. So
for example, I can talk to you, and I can
also do machine learning research, and I can paint, and
I can do a lot of other things. And I
think this is what I mean by generali, And we
do have all of these amazing hints from nature of
what we need, at least at the high level, so
we know that we need some prior information and built
(32:33):
in the model, but also really strong adaptation, especially early
in life. Humans are really good at this. This is
what makes us so unique that we are able to adapt.
But we're also really able to prune unneeded information, and
this is becoming also important to think that, oh, maybe
in our models we also need to do that to
prune connections and it makes connections to sparse models.
Speaker 3 (32:54):
And so on.
Speaker 4 (32:55):
But also we have memory systems and the importance of
building future predictions. So all of these are useful hints
that we can then try to ask, Okay, how do
we implement this using our computational back ends, because what
we have is a different computational back end than the
brain has. That is very clear. But maybe we can
(33:18):
use all of these hints to build these models. And
I think these hints, together with the process, the progress
that we've made in recent years, makes me think that
we will be able to build such flexible enough models
that such that they can solve multiple tasks and they
can build a like general intelligence. But one thing I
(33:42):
also think is that we have to be really open
minded about these computational back ends that we will have
to use to get there. But I do think humans
as a species will be able to build generally at
some point a.
Speaker 2 (33:56):
Few thoughts because you mentioned very interesting things. So based
on what you said, I guess all the advancements in
understanding the human brain, even euroscience, are key when it
comes to pursuing general AI. Right, that's one of the things.
Do you somehow also work on this in the mind
(34:18):
you have people working on this?
Speaker 4 (34:20):
Yeah, Yeah, the mine has folks working on neuroscience, and
I think even myself, even though I don't work directly
on neuroscience, I try to keep an eye on parts
of the field, and I'm generally interested because I think, yes,
I said, I think it provides a very useful guiding
(34:40):
example of the high level of what we will need
to I mean, intelligent animals are the only examples of
intelligence that we have, so it makes sense to get
some hints from what we need from there, It really does.
Speaker 2 (34:58):
You mentioned a word that really makes sense, being adaptable.
That's one of the key strengths that humans have, that
we are able to respond in a correct way sometimes
to new stimulus or to new situations in scenarios, even
if we don't have the require data set from before.
(35:21):
Talking in algorithm terms, but would you expect that an
agent would so.
Speaker 4 (35:30):
Not?
Speaker 2 (35:31):
Adaptivity but creativity to some problems or would it be
fake creativity? Can can you create real creativity in an agent?
What do you think?
Speaker 3 (35:42):
I mean?
Speaker 4 (35:42):
I guess it defends on what you define creativity as.
But I think one of the components that we would
miss and a lot of what you said is kind
of adapting to a new situation as counterfactual reasoning, which
is something that humans can do very well, and in
some sense it's only at the beginning of really being
(36:05):
integrated into machine learning systems. So I think that's one
step in that direction.
Speaker 2 (36:11):
Anyhow, we're still far from that. I think there is
a lot of work to be done on that aspect.
But what do you think are the main challenges today?
Because I've wrote a couple of articles about responsible AI,
why in our field, I think it's really important that
we take into account that we shouldn't be or that
(36:32):
we should be eliminating biases in some complex predictive models.
Imagine that as we work with an insurance company and
we need to help them with I don't know, a
credit a score or a health score for them to
be able to provide a correct assessment on the pricing
sort of say, right, all of these needs to be
(36:52):
quite open and honest and clear when it comes to
how the algorithms work. So apart from RESPONSIBILII and you
can talk about responsibility as well. What are the main
challenges today, Because you know, computational power advanced quite a lot.
I don't think that's the Showers stopper right now. What
(37:15):
do you think are the challenges that we are facing
in our niche or our field of action at the moment.
Speaker 3 (37:22):
Yeah, I think what you said is spot on.
Speaker 4 (37:24):
I think if we want to deploy machine learning safely
and at scale, we really have to know the effects
of its biases and minimize them, because I think we
need the benefits of machine learning to be equally experienced
in society, and we really have to be careful not
to perpetrate or exacerbate even existing biases or stereotype. And
(37:48):
it's not only that the training data can have biases,
because if we get data from humans and train a
model on that, then it will have it can have
some of our existing biases. But models can introduce biases themselves,
and I think it's important to know that. And machine
learning has the promise to be a great equalizer. So,
(38:09):
for example, a model that we train from data from
the best doctors in the world has the promise to
provide good diagnosis to those that don't have access to
the best doctors in the world. But that's only if
this model performs well on data from all of these
people that were going to apply the model on. And
(38:30):
this is not something that we should take for granted
that it will generalize in that way. Absolutely not. So
this is something that I think needs to be thoroughly
checked and investigated before the model gets deployed. And also
on the research side, I think we've seen in the
last few years a lot of work and I hope
(38:52):
that we will see more on how to make sure
that we train models that do not produce or acacerbate biases,
but also that we can have methods to basically to
basically check for this.
Speaker 2 (39:06):
Do you think that at some point we will have
some sort of institution or organ as part of the
government that will have I don't know, another in process
in place for this or yeah.
Speaker 4 (39:19):
I think it might also depend on the branch in
which the model gets applied. So obviously, if you apply
a model in medicine, there are different Yeah, there are
different effects that one can have and as you said,
also an insurance. It's very important to make sure that
(39:40):
everyone gets a fair assessment. So I think it might
depend on the exact industry.
Speaker 2 (39:47):
Okay, I don't know how it is for you, but
data science and AI and both of the fields analytics,
if you may say, or advanced analytics are getting a
lot of high around it. And you know, some of
my friends they tell me you're working on what's the
future of humanity somehow and probably you're part of well
(40:09):
you're part of the mind, so you probably get a
lot of these too, do you.
Speaker 3 (40:16):
Yeah.
Speaker 4 (40:16):
I think people generally are quite interested more in the
little less recent years, and they know when I say
I work on machine learning or II, they know about
what it is about and they have opinions on it.
And a lot of people also know about deep mind
for example. So yeah, I think this has changed in
(40:37):
the last five to ten years. Absolutely.
Speaker 2 (40:40):
Why do you think it has changed that much? Is
it because of the documentaries? Is it because of I
don't know, companies analytica, with the with the elections. What
do you think has created this some turmoil in a
positive way? Right? This this awareness of this field.
Speaker 3 (41:03):
Yeah, that's a good question.
Speaker 4 (41:04):
I mean, I think one thing that has happened since
the twenty tens, for example, is that machine learning has
really gone from this interesting theoretical endeavor to a practical
field which can make a big difference in products, especially
now when we have high dimensional, large data. And I
think this is what neural networks and deep learning makes
(41:27):
do differently. And I think the primary example of that
earlier on was the huge increase in performance on image NEET,
which is this hard image classification benchmark.
Speaker 3 (41:39):
And I think that showed that.
Speaker 4 (41:40):
The recipe of okay, deep learning, so you take neural
networks plus more computation plus more data, can work very
well for these high dimensional, very difficult problems. And since
then we've seen other ones such as GPT, which you've
just mentioned, and now more recently Alpha folden science, and
I think with people see this big proof of concepts
(42:02):
of one machine learning can do in these large settings
that were very hard to do before, they are naturally
attracted to it. And then we see people either in
industry or in other sciences trying to tackle problems using
machine learning.
Speaker 2 (42:20):
Interesting. Okay, just before we close the goal off, I
wanted to appreciate you taking the time to take this call.
I'm really learning a lot and I hope you're enjoying
this cull too. But before we close the goal off,
could you please recommend me someone to join me or
join us in the next episode of Data Stand Up.
Speaker 3 (42:43):
Yeah.
Speaker 4 (42:43):
So, I think it's always great to chat with people
from various backgrounds. So I think you can maybe reach
out to someone from academia or someone working in other
countries that the guests that you've had so far we're
not working in. So, for example, I know a few
folks doing PhDs and machine learning in Romania and I
might be able to connect you with them if that's
(43:04):
something that you're interested in.
Speaker 2 (43:06):
Yeah, of course. I mean sometimes when you when you
speak with people that are already working in a specific company,
they have those biases of being exposed to a set
of machine learning or dat analytics applications within that industry
and within that company, and there is so much more
(43:27):
to that that we want to learn and we want
to discuss about. So if they are working on new
application and they are expanding this field, we want to know,
we want to know about that, Yeah, of course. And
another question, because you did say that you read a
lot sometimes papers, whatever it is. But give me a book,
(43:52):
a newsletter or something that you sometimes read that you
highly recommend to stay up to the latest in this field,
or maybe because it's a book that's entertaining, and that's
it's really a learning tool for anyone that I know
is trained to have started in something that relates to
(44:14):
a YI or machine learning.
Speaker 4 (44:15):
Yes, as a textbook, I really really like Kevin Murphy's
Machine Learning A Probabilistic Perspective and I always like having
it on my book, so that that's a recommendation, And
he has version two coming up now, so I think
that's that's something to look out for. But I think
for recent work fair and Khosar has an excellent machine
(44:37):
learning blog post at Inference dot VC, and if you're
following the literature closely, I think this is an interesting
blog post to look out for.
Speaker 3 (44:51):
I think it's it's it's a really great resource.
Speaker 2 (44:53):
Okay, fantastic, Thank you so much, Mihaila. I really enjoyed
making this call, and again I appreciate the time that
you took to discuss with this all of all of
these advanstment advancements in in AI. Thank you again, and
I hope you you have a really nice day.
Speaker 3 (45:13):
Thank you so much for having me.
Speaker 2 (45:15):
Thank you Mihailer, have a good one. Bye bye Mu
Bedroom is an independent data science and aim specialized in
data driven business change. In this podcast, our guests help
us spread knowledge and experience with our listeners.
Speaker 2 (00:29):
Good morning, Mihaila, how are you doing today?
Speaker 3 (00:32):
Good morning, I'm great. How are you?
Speaker 2 (00:35):
I'm fine. I believe that the autumn is officially here.
It's getting older than ever here in the earth of Spain.
How is it in London?
Speaker 3 (00:45):
It's good.
Speaker 4 (00:45):
It's actually I was looking outside before the recording it's
quite sunny. But I also feel like, yeah, the summer
has ended. I think Saturday we had our last real
summer day.
Speaker 3 (00:56):
It was really warm and sunny.
Speaker 4 (00:58):
But I think it's yeah, we're we're getting braced for
autumn and winter.
Speaker 2 (01:04):
Are you a summer person or more like you like
winter and the Christmas lights? What do you prefer?
Speaker 3 (01:10):
I am a huge summer person. I love summer.
Speaker 2 (01:13):
Me too.
Speaker 4 (01:15):
I yeah, I really don't like winter, So for me,
this is a bit sad summer is over.
Speaker 2 (01:24):
Same here, okay, So tell me what do you have
in the agenda in the schedule for today? Is it
packed with meetings?
Speaker 3 (01:34):
Nos?
Speaker 4 (01:34):
I mean I always start my day with a green tea,
so I have it here with me and after we
talk today, I'll have some time for a little bit
of what I like to call focus time, so I'll
be thinking a bit about my next steps research wise,
and then i'll yeah, I'll have one or two meetings,
but yeah, not too much today. I think in general,
(01:56):
it does vary a lot from time to time where
they are. I'm sorry, I just spent my whole day
focused on paperwriting or doing some proofs. But now it's
a little bit of that period of trying to yeah,
reassemble and figure out what to do next.
Speaker 2 (02:13):
So whoever is listening to this here or see, we'll
hear about papers like you just mentioned. So I think
it's appropriate to provide some context on who you are,
where you come from. So youar, Mihaila, you're currently in
Deep Mind. You are working as a research I wouldn't
say engineer, right, how would you describe it? What's the
(02:35):
what's the what's the role?
Speaker 3 (02:36):
Yeah?
Speaker 4 (02:36):
My role is as a research engineer, that's right. So
that's I'm a stuff research engineer and I've been working
at deep Mine for yeah, five years in this strole
and at the same time, I'm also doing a PhD
at UCL.
Speaker 2 (02:51):
Thank you, And we have you here because we did
a podcast recording also with Cercio, with Sergio as you
may call him, who who works in deep Mine in
London and who was also part of Data Stand Up
Spanish podcast earlier this year. Right, he was the one
(03:11):
who told you to take part in this.
Speaker 3 (03:13):
That's right, that's right.
Speaker 4 (03:14):
He's an amazing colleague and a really nice person as also.
I was really happy when he reached out and asked
me what I want to do this and.
Speaker 2 (03:25):
We're happy to have you here. Okay, so you completed
your bachelor and your master's in London. I guess something
that's interesting to anything, to anyone listening to this conversation
would be what did prepare you well for your current role?
Because you're in a deep tech and a research firm
(03:46):
like deep Mind, And what did you miss maybe theoretical
knowledge or any sort of a practical experience before joining.
But yeah, what did prepare you well? And what didn't
sort of say?
Speaker 4 (04:01):
Yeah, this is a very good question. I think my
degree really focused on providing us with the good technical
expertise and also good knowledge on software engineering, but also
we had a lot of advanced courses on mathematics and
machine learning.
Speaker 3 (04:14):
Which had been extremely useful.
Speaker 4 (04:16):
So I think on the technical side, I was very
well equipped after finishing my degree. But I think one
aspect of research that is really hard to learn unless.
Speaker 3 (04:26):
You're really doing it.
Speaker 4 (04:28):
So it's kind of hard to teach us prioritizing a
lot of the research ideas that you or the team
might have. I think maybe earlier in life, I thought
all research ideas are really hard to get, and then
once you have one, you just spend your whole time
pursuing that. But in practice, we often have many research
ideas and a lot of the hard part of my
(04:50):
job is to also try to assess, okay, which are
the right research ideas to chase, and that depends on
risk level for that idea, but also or how promising
the impact of that idea is. So this is something
that you always kind of constantly have to balance and
it's a bit hard to teach. And the skill that
(05:10):
I found that I had to really learn on the
job and become better and better at it in time, and.
Speaker 2 (05:17):
So that balance is somehow support it. I mean that
decision that has to be made on what you focus
on is somehow supported by alphabet because for the listeners
that do not know, which I guess would be only
a few. But the mind is the AI and research
unit of WILL, which ultimately belongs to Alphabet. But that's
(05:41):
the business as we will not imposed but guide you
through the different fields that would be of interest for
you to pursue or do our research on something that's
your interest. How does it work, Well, we do.
Speaker 4 (05:55):
Have a lot of research freedom, so I think it
might operate relatively independently. But of course it's a matter
I find of trying to also convince your colleagues that
if you have an idea, that's worth pursuing, because often
you might have an idea. But again coming back to
(06:16):
what I've previously said, there are many ideas. So I
don't really work and avoid I work with my colleagues,
and it's a matter of together between all of us
trying to figure out what's the right thing to focus
on right now.
Speaker 2 (06:30):
So you're completing your PhD at the moment, right what's
the field of a study?
Speaker 4 (06:34):
Yeah, right now I'm in my PhD. I'm generally focusing
on learning probability distributions, and I'm really interested in learning
probability distributions and learning them efficiently. Because this is really
at the core of machine learning. So classification, generative modeling,
reinforcement learning, they all rely on learning distributions efficiently. In classification,
(06:56):
well learn a conditional distribution label given the data. In
generative modeling we might learn an unconditional or also conditional distribution,
and in URL in reinforcement learning, you learn a policy
which is also distribution condition on the state of what
action you should do next. So this is why I
find learning probability distribution is extremely important because if you
(07:20):
make progress here, you can make progress in a lot
of the fields. And I'm focusing a lot on what
kind of loss functions one should use to learn these distributions,
but also most recently a also how to optimize and
regularize neural networks that are learned to represent these probability distributions.
Speaker 2 (07:40):
I'm curious, then, so you've mentioned lost functions. I guess
that your work, as it's part of a research exercise,
it goes beyond yes, coding and creating a library that
summarizes this for prefers, this for any other inducer. But
it's about doing that mathematical research, not on paper, but
(08:02):
you have to do that deep work on what you
change in the back end sort of say right.
Speaker 3 (08:08):
Absolutely, absolutely. Yeah.
Speaker 4 (08:09):
So for example, in one of our recent papers, we
use mathematics, as you've just said, to try to understand
what gradient descent actually does in two player games, and
we can quantify that. For example, in our case, we
use an ode or ordinary differential equation, and based on
the math, we then see, oh, we can treat this
(08:30):
algorithm that then improves the performance.
Speaker 3 (08:33):
But yeah, there's a lot of new math.
Speaker 4 (08:36):
Let's say that describes a particular situation. In this case,
it was great and descent with games. And then we
take that math and we do something with it. And
the something often involves coding an algorithm so that we
can test our mathematical predictions empirically.
Speaker 2 (08:54):
Okay, we can talk about this later because I'd be
curious to know if these pre col demonstrations also go
to the business side of Google, for instance. Okay, back
to a few years back. I'm curious, what was the
Mind working when you joined back in twenty sixteen.
Speaker 4 (09:12):
Yeah, I mean the Mind was working on a lot
of really interesting things and a lot of the core
methods we still use today.
Speaker 3 (09:18):
So there was a lot of work.
Speaker 4 (09:19):
On reinforcement learning, generative models and also enhancing, for example,
reinforcement learning methods with memory retrieval and neural systems. And
I think, what's what's interesting when I saw this question
is looking back, how a lot of these core methods
are still really used today in our models.
Speaker 2 (09:39):
And so he told me, you're part of a research
team now trying to look at the big picture. And
probably you're familiar with the structure of the organization. How
many units is the mind composed of? Somehow what is
each of these ones doing?
Speaker 4 (09:59):
Yeah, so I am part of the research team, and
deep Mind also has an apply team, which focuses a
lot on bringing deep Mind research a lot of the
research that I talked about before to the world into products,
for example, and.
Speaker 2 (10:15):
If there was a need to implement those algorithms that
you've developed and you've tested into a product that I
don't know is part of the Google Cloud platform suite,
that be this apply team, right.
Speaker 3 (10:29):
Yeah, it depends.
Speaker 4 (10:31):
It can also be shared, so for example, the research team,
if they come up with an idea, they can also
collaborate of course with the apply team.
Speaker 3 (10:39):
Everything.
Speaker 4 (10:41):
One thing that I really like about depind is that
everything is flexible, so you can be part of the
research team, but if you have come up with an
idea and you really want to see that into a product,
you can still work on that.
Speaker 2 (10:53):
And when you joined, how many were you? I'm just
asking because I started to know about deep Mind when
the famous documentary of Alpha Go appeared in my Netflix
recommendation list and I watched it. I loved it, actually,
so I'm curious to know how did the company grow
(11:14):
from when you joined till today? How did you leave
this growth?
Speaker 4 (11:21):
Yeah, I mean I think I don't remember exactly how
many we were, but certainly there has been a lot
of growth scenes and I think the what that allows
the company to do, and I think what you can
see is focusing on more research directions and more very
(11:42):
ambitious products and projects. So, for example, as you've seen recently,
there was the Alpha fold result where a deep Mind
was able to produce a machine learning method which is
very good at protein folding and really advancing the state
of the art and science. So I think this this
is primarily what growth has achieved. I think the culture
(12:05):
and the atmosphere is very similar to the one that
I experienced when I joined in two thousand.
Speaker 2 (12:11):
Sixty, which I guess is one of the hardest things
to keep right. I mean, when you grow that much,
you add different people to the team, making sure that
those values, those initial values stay, that's really important as well.
Okay for anyone listening to this conversation that maybe feels
ambitious about joining deep Mind because he feels or she
(12:34):
feels that he wants to be a part of that
research mission that you're mentioning because that's her interest. What
do you look for in candidates? Have you ever take
part in hiring selection processing, and even if you didn't,
what do you think are the best traits to be
a good contributor to the Mind's work.
Speaker 4 (12:53):
Yeah, so I have been part of the hiring process
at various stages. I think the exact requirements will depend
on the role. So there are different technical interviews for
software engineering role, for a research engineer or for a
research scientist. But I think in general we are looking
for candidates who are just interested in artificial intelligence and
really like solving different problems because that's a lot of
(13:16):
the day to day of what we do, and also
are keen of being part of the defined team.
Speaker 2 (13:23):
And I guess that one of the most important things
when you are part of this cutting edge field is
to be up to date, read a lot, make sure
that you pay attention to everything that it's being published.
Do you keep time in your agenda in your schedule
(13:43):
to go through some of these papers. How much time
of your time relates to reading and consuming information from others.
Speaker 4 (13:51):
Yeah, I do spend quite a lot of time reading
and consuming papers, and I think it's really important to
stay connected with the research ideas from the community. I
also really like to take a few days just once
in a while to just read, so basically to block
out my calendars or to speak and just say, okay,
I'm going to now immerse myself into what's going on
(14:13):
either and through something that I'm researching right now or
something that is maybe somewhat related and I'm trying to
learn about. And I think, yeah, it's quite important for me.
And I think a lot of ideas and inspiration comes
from us. I mean, what's going on in the literature.
Speaker 2 (14:29):
I completely agree. I mean some of the summer when
I was taking some time off, I was going through
some of the articles that you can find in platforms
such as medium. When you select machine learning, AI or
data science, you can find some of the I mean,
those aren't as deep a paper, but you can find
(14:51):
some applications that provide you with ideas. And I believe
that taking some time off to reflect on what you're
doing us that we do. I mean, part of our
work is research. Even if we are a putique practitioner
of data science and AI for our clients, we still
need to stay ahead of what the competition is doing
(15:14):
and to make sure that we provide what's the latest
advice advancements in technology and applications. So I guess that
reading and consuming information is critical for professionals, regardless of
you know, being research or not. Okay, so you work
in a research field that, as I said before, is
(15:35):
categorized as categorized as cutting edge, but you get to
see these developments applied in real life sort of say,
because you've mentioned alpha fold, but did you ever encounter
a technology that you use in your day to day
that you think, Okay, part of the work that we
(15:56):
did in deep Mind is somehow embedded into the device
or this application, and how does it feel? Yeah?
Speaker 4 (16:04):
Absolutely, I think there are many examples of cutting edge
research which are now part of alphabet technology. So for example,
wave Net, which was a research project, is now part
of a lot of the text to speech products at Google.
And another nice recent example is traffic prediction as part
of Google Maps using graph nets. So there are many
(16:28):
others as well. So I think this is also what
makes machine learning research so appealing, is that you can
have these research ideas and actually they do lend themselves
to being very useful in practice. Then they can be
used by millions of millions of people.
Speaker 2 (16:46):
So when I'm using Google Maps, I should think of
the mind developing those predictions of what's the traffic going
to be like in my city? Correct?
Speaker 4 (16:58):
Yeah, yeah, it's part of the prediction and the recent
traffic Yeah, amazing.
Speaker 2 (17:05):
The thing is how how does the algorithm work in
that sense, because I guess that the system is being
fed continuously with the traffic information from all of the
different cities around the world. Is that information processed within
your device or how does it happen?
Speaker 4 (17:20):
So I'm not aware of how this actually works on
the on the application level. Okay, So yeah, I don't
know exactly how does is implemented in the application.
Speaker 2 (17:31):
Okay, So you've mentioned alpha fault before, and I also
mentioned that the mind somehow I started to become very
widespread and everyone started to knew about Alpha Go when
it was released in Netflix because of your developments for
(17:53):
the Alpha Go game, for the Go game. Sorry, so,
can you tell me something about this that we may
not know in terms of I don't know developments. I
wouldn't say secrets, but probably I don't know anecdotes about
the development of Alpha Go software or the Alpha Hool
(18:14):
development that could be of interest to the listeners.
Speaker 4 (18:18):
Well, yeah, I don't know if there's anything. I think
one thing that is really nice about these projects is
seeing kind of the excitement around the office, and I
think that's very hard to replicate or to understand. I
think I joined in Mind twenty sixteen, so the lisad
all Alpha Go matches had already happened, but there were
some further matches in twenty seventeen, and there was so
(18:41):
much excitement in the office around these matches, and I
think that really, yeah, it really has an impact on
your day to day to kind of come to work
and to see all of that unfolding life. And I
think it's also similarly the case with Alpha fail to
know that this amazing groundbreaking research is happening, and to
(19:03):
have chats with people about it and all of that
really keeps you very, very excited.
Speaker 2 (19:08):
Something that amazed me in regards to that when I
watched that documentary is that your colleagues were excited but
at the same time a little bit afraid that the
software that they developed, those algorithms, we're going to surpass
(19:28):
the human capabilities of famous players and very capable players,
And they were excited about their development being successful, but
at the same time they were feeling bad about the
human on the other side. I'm not sure if if
you saw that.
Speaker 4 (19:46):
Yeah, I mean, it's always interesting, right because we are
all humans. It's yeah, it's a bit of it is
an interesting human experience to see that you you're able
to build algorthon, which is very good, which is of
course promising because we're hoping to apply all of these
(20:11):
machine learning advances to make the world a better place,
which is obviously quite quite exciting. But also on the
other hand, as you said, it's hard not to have
empathy for the human who is the best, let's say,
go player in the world. So yeah, it's a bit
it's an emotional mix, but I think still nonetheless quite
(20:33):
quite an exciting experience to have.
Speaker 2 (20:36):
Yes, And related to that, well, something that we do
ourselves studies somehow related to that is when we start
to work with an organization where we have to reinforce
or promote data culture, where we are developing AI mechanisms
for the humans at that company to be used, is
(20:57):
that these mechanisms hurt created, designed and implemented to empower
the people at that organization and to not substitute them,
because that's sometimes a mix of feelings that people have,
right like, they see AI as something that could be
(21:17):
evil and substitute in some of the capabilities that these
guys have or you know that of the things they
are doing in their day to day, But it's more
about empowering those individuals to make the world a better place,
if you want to see it that way. That's how
we think and that's why we try to promote. But
(21:38):
at the same time, yes, of course, when you create
something that has strong capabilities and probably can make decisions
faster than a human, you have this feeling that is
overwhelming sure to say. Okay, so tell me about some
of your achievements in research or even in general, how
(22:00):
they impact it your life. On one side, and maybe
your work in alphabet on the other, in the mind
as part of alf IT on the other.
Speaker 4 (22:12):
Yeah, I mean what really drives my research is this
motivation that I personally have to understand what's going on.
So a lot of what I've done is this kind
of research that focuses understanding the mechanism by which certain
methods work or don't work. So, for example, we might
know that there are certain limitations of sub types of methods,
but we don't really understand why it fails in that way.
(22:34):
So I'm really interested in understanding that, and I've done
that for a ganzo, a type of generative model that
learns to generate, for example, images using a two player games,
but also variational inference and stochastic gradient estimation, and more recently,
as I mentioned, optimization in games and regularization in neural networks.
So that's the understanding part. And then once you have
(22:57):
this understanding part, you can really use that. Either myself
or other colleagues have done that taken that understanding to
improve these methods in parts of either gans or variational
inprints and so on. So this is on the research
side and on the coding side. I also been part
of designing research libraries that have been used and defined
or more widely at alphabet.
Speaker 2 (23:17):
Okay, when you talk about regularization in games, does it
relate to I don't know about or a decision that
has to be made but was never made before is
somehow correctly done. Can you tell me more about this,
explain it a little bit more.
Speaker 4 (23:35):
Yeah, So here I met to regularization the context of
neural networks. So often when we played these games, we
still implement our two players these games as neural networks.
So in the case of gans, for example, you have
a discriminator which is going to be a neural network,
and a generator which is going to be a neural network.
And often when you train these models, we see that
there are certain methods that basically regularize the neural network space.
(24:00):
The functional space that represents the discriminator or de gerator
can be very useful at improving performance. But sometimes we
know why and sometimes we don't. So for example, when
I say regularization, I mean sometimes let's say L two
regularization when you add this term to the lost function too,
(24:21):
let's say, make sure to penalize the L to norm
of the way. That's a type of regularization, But these
days we might have more complicated approaches or more complex
approaches to regularize these models. And for example, one that
I was particularly interested in is smoothness with respect to inputs,
So that says, I want on your own network that
(24:42):
if I change my input a little bit, my output
doesn't change too much, which is kind of an intuitive
thing to want from your model.
Speaker 2 (24:51):
That's really interesting. Got you okay? And living aside all
that technical research works sort of say, I think last
year and the previous one was a bit of a
weird here in regards to the pandemic and so on
and so forth. How did this impact your work? What
did you learn? How are you utilizing these learnings?
Speaker 4 (25:14):
Yeah, this is a very good question. I mean, of
course it affected everyone. We went proworking in the office
and seeing everyone every day to working from home. And
also I think what the pandemic has particularly done for
me is reminded me how important personal circumstances are, even
(25:34):
at work and for work, and how we really have
to be mindful that our colleagues have different personal circumstances
and challenges. And I think this was really made obvious
by the pandemic. Some people were alone at home, which
was obviously very hard. Some people were struggling because they
had to work and take care of their children, and
(25:55):
there are many made there's the variety of experience of
how this was felt was quite large, and I think
it really made us more mindful that even though maybe
one person is doing well, another person is doing not
so well for different reasons. And I think we really
have to account for that in our day to day
(26:15):
work and support each other. And I'm planning to stay
a lot more mindful about this as the COVID situation
hopefully gradually improves.
Speaker 2 (26:26):
So it's about feeling and showing that empathy to co worker,
colleagues and even friends. Right, Like, you don't know what
the other person may be going through or feeling at
a specific point in time, so you just need to
be with her with him, right.
Speaker 4 (26:42):
Absolutely, absolutely, And I think sometimes what I've also noticed
is that because of the pandemic, maybe people are more
likely to share that something is going on or the
situation is hard, which makes it easier for example, to
be there for them and to help. And then I'm
I'm hoping that we will continue having that that culture
(27:05):
of sharing and saying actually things are not so good
right now, so maybe we can.
Speaker 3 (27:10):
Yeah, we can adjust our workload or.
Speaker 4 (27:13):
Anything that we need. Sometimes just talking to someone can
can really help as well.
Speaker 2 (27:19):
I'm curious. I didn't think of this till now, but
did you, as you know, deep Mind, work on anything
in regards to fighting off the pandemic or its effect
for COVID? Was it any Was there any development in
regards to that.
Speaker 4 (27:37):
I don't really know myself, so I I haven't been
involved in in this directly.
Speaker 2 (27:46):
Yeah, I didn't know either. I mean, this was just
out of the blue. As a curiosity. We did something
ourselves that wasn't deep dech, but it was hard because
we well, i'll explain to you what we did. We
created uh recommendation engine for people that wanted to travel.
So depending on the location where they wanted to travel,
(28:09):
we would be retrieving and gathering data from different sources,
some in regards to COVID, some in regards to other
security aspects of that location, and providing the end user
or the person with a recommendation index about how safe
it was to travel to that location. It wasn't only
related to COVID that I think it was more related
(28:31):
to rebuilding and regaining confidence of the people you know
that needed to travel or that had to travel. And
also because I think or we thought this could have
an impact on the economy, we called this this tool
check check doto. And this is not me doing any
(28:51):
promotion here of the tool, because this was something that
we did provano sort of say, just because we felt
it could have an impact after the after COVID nineteen.
I completely assume that probably I don't know, the Mind
did some sort of COVID prediction in regards to how
it was going to be I don't know, moving around
(29:16):
London or the UK, I don't know. Okay. Now, living
aside your work at deep Mind and yourself even do
you also follow on pay attention to your biggest allies
or the say, in terms of making progress in the
field of AI, which probably is open AI and their
(29:39):
advancements such as CBD three and other developments. Because you've
mentioned that you were talking about text generation before, I
think and somehow it's related to this, so probably this
is of your interest as well, right DBD three.
Speaker 4 (29:55):
Yeah, I mean I do follow advances in the field,
generally both from academia and other result slabs and so
of course I have read and been extremely impressed by
the results of GIPD three.
Speaker 2 (30:06):
Do you do you expect that at some point as
as humans will be able to communicate completely openly and
fluently with computers, because I know that the hardest thing
here for computer is to improvise right, generalize and be
able to make decisions on the goal in the in
the in the worlds to be chosen and the words
(30:28):
to be used. And how far do you think we
are from this? Because one thing is you know, text
generation and chatbots. But at some point, I guess we
expect a complete fluent conversation with a human, even discussing
even feelings. You think we will get to some point
We will get at some point of this.
Speaker 4 (30:48):
Yeah, I mean I am one of the people that
actually believes that we will reach general AI, and I
think if we think every reach General II, then we
will also have a good converse way to converse with
these agents that are acting in the world alongside us.
Speaker 2 (31:10):
Okay, so you're a film believer that we will reach
general AI. So humans that can feel emotions, that are
able to generalize in problem solving, in movements, in everything,
do you think this will exist at some point, So.
Speaker 4 (31:24):
It's not humans agents. I mean I think I think so. Yeah,
I think we will. I'm not, let's say, one hundred
percent sure, but I think I'm generally thinking towards us
quite confidently, because I think so. Our brains are incredible
computation machines, and they have been able to produce high
(31:48):
levels of intelligence. And while our computational models that we
use right now in our computers are different than those
in the brain, the really nice thing is that we
actually get some really useful hints from the brain of
what we need to do to build these general agents
which can solve multiple tasks. So when I mean generali,
I don't necessarily think about feelings necessarily, but an agent
(32:11):
that like us can solve a variety of tasks. So
for example, I can talk to you, and I can
also do machine learning research, and I can paint, and
I can do a lot of other things. And I
think this is what I mean by generali, And we
do have all of these amazing hints from nature of
what we need, at least at the high level, so
we know that we need some prior information and built
(32:33):
in the model, but also really strong adaptation, especially early
in life. Humans are really good at this. This is
what makes us so unique that we are able to adapt.
But we're also really able to prune unneeded information, and
this is becoming also important to think that, oh, maybe
in our models we also need to do that to
prune connections and it makes connections to sparse models.
Speaker 3 (32:54):
And so on.
Speaker 4 (32:55):
But also we have memory systems and the importance of
building future predictions. So all of these are useful hints
that we can then try to ask, Okay, how do
we implement this using our computational back ends, because what
we have is a different computational back end than the
brain has. That is very clear. But maybe we can
(33:18):
use all of these hints to build these models. And
I think these hints, together with the process, the progress
that we've made in recent years, makes me think that
we will be able to build such flexible enough models
that such that they can solve multiple tasks and they
can build a like general intelligence. But one thing I
(33:42):
also think is that we have to be really open
minded about these computational back ends that we will have
to use to get there. But I do think humans
as a species will be able to build generally at
some point a.
Speaker 2 (33:56):
Few thoughts because you mentioned very interesting things. So based
on what you said, I guess all the advancements in
understanding the human brain, even euroscience, are key when it
comes to pursuing general AI. Right, that's one of the things.
Do you somehow also work on this in the mind
(34:18):
you have people working on this?
Speaker 4 (34:20):
Yeah, Yeah, the mine has folks working on neuroscience, and
I think even myself, even though I don't work directly
on neuroscience, I try to keep an eye on parts
of the field, and I'm generally interested because I think, yes,
I said, I think it provides a very useful guiding
(34:40):
example of the high level of what we will need
to I mean, intelligent animals are the only examples of
intelligence that we have, so it makes sense to get
some hints from what we need from there, It really does.
Speaker 2 (34:58):
You mentioned a word that really makes sense, being adaptable.
That's one of the key strengths that humans have, that
we are able to respond in a correct way sometimes
to new stimulus or to new situations in scenarios, even
if we don't have the require data set from before.
(35:21):
Talking in algorithm terms, but would you expect that an
agent would so.
Speaker 4 (35:30):
Not?
Speaker 2 (35:31):
Adaptivity but creativity to some problems or would it be
fake creativity? Can can you create real creativity in an agent?
What do you think?
Speaker 3 (35:42):
I mean?
Speaker 4 (35:42):
I guess it defends on what you define creativity as.
But I think one of the components that we would
miss and a lot of what you said is kind
of adapting to a new situation as counterfactual reasoning, which
is something that humans can do very well, and in
some sense it's only at the beginning of really being
(36:05):
integrated into machine learning systems. So I think that's one
step in that direction.
Speaker 2 (36:11):
Anyhow, we're still far from that. I think there is
a lot of work to be done on that aspect.
But what do you think are the main challenges today?
Because I've wrote a couple of articles about responsible AI,
why in our field, I think it's really important that
we take into account that we shouldn't be or that
(36:32):
we should be eliminating biases in some complex predictive models.
Imagine that as we work with an insurance company and
we need to help them with I don't know, a
credit a score or a health score for them to
be able to provide a correct assessment on the pricing
sort of say, right, all of these needs to be
(36:52):
quite open and honest and clear when it comes to
how the algorithms work. So apart from RESPONSIBILII and you
can talk about responsibility as well. What are the main
challenges today, Because you know, computational power advanced quite a lot.
I don't think that's the Showers stopper right now. What
(37:15):
do you think are the challenges that we are facing
in our niche or our field of action at the moment.
Speaker 3 (37:22):
Yeah, I think what you said is spot on.
Speaker 4 (37:24):
I think if we want to deploy machine learning safely
and at scale, we really have to know the effects
of its biases and minimize them, because I think we
need the benefits of machine learning to be equally experienced
in society, and we really have to be careful not
to perpetrate or exacerbate even existing biases or stereotype. And
(37:48):
it's not only that the training data can have biases,
because if we get data from humans and train a
model on that, then it will have it can have
some of our existing biases. But models can introduce biases themselves,
and I think it's important to know that. And machine
learning has the promise to be a great equalizer. So,
(38:09):
for example, a model that we train from data from
the best doctors in the world has the promise to
provide good diagnosis to those that don't have access to
the best doctors in the world. But that's only if
this model performs well on data from all of these
people that were going to apply the model on. And
(38:30):
this is not something that we should take for granted
that it will generalize in that way. Absolutely not. So
this is something that I think needs to be thoroughly
checked and investigated before the model gets deployed. And also
on the research side, I think we've seen in the
last few years a lot of work and I hope
(38:52):
that we will see more on how to make sure
that we train models that do not produce or acacerbate biases,
but also that we can have methods to basically to
basically check for this.
Speaker 2 (39:06):
Do you think that at some point we will have
some sort of institution or organ as part of the
government that will have I don't know, another in process
in place for this or yeah.
Speaker 4 (39:19):
I think it might also depend on the branch in
which the model gets applied. So obviously, if you apply
a model in medicine, there are different Yeah, there are
different effects that one can have and as you said,
also an insurance. It's very important to make sure that
(39:40):
everyone gets a fair assessment. So I think it might
depend on the exact industry.
Speaker 2 (39:47):
Okay, I don't know how it is for you, but
data science and AI and both of the fields analytics,
if you may say, or advanced analytics are getting a
lot of high around it. And you know, some of
my friends they tell me you're working on what's the
future of humanity somehow and probably you're part of well
(40:09):
you're part of the mind, so you probably get a
lot of these too, do you.
Speaker 3 (40:16):
Yeah.
Speaker 4 (40:16):
I think people generally are quite interested more in the
little less recent years, and they know when I say
I work on machine learning or II, they know about
what it is about and they have opinions on it.
And a lot of people also know about deep mind
for example. So yeah, I think this has changed in
(40:37):
the last five to ten years. Absolutely.
Speaker 2 (40:40):
Why do you think it has changed that much? Is
it because of the documentaries? Is it because of I
don't know, companies analytica, with the with the elections. What
do you think has created this some turmoil in a
positive way? Right? This this awareness of this field.
Speaker 3 (41:03):
Yeah, that's a good question.
Speaker 4 (41:04):
I mean, I think one thing that has happened since
the twenty tens, for example, is that machine learning has
really gone from this interesting theoretical endeavor to a practical
field which can make a big difference in products, especially
now when we have high dimensional, large data. And I
think this is what neural networks and deep learning makes
(41:27):
do differently. And I think the primary example of that
earlier on was the huge increase in performance on image NEET,
which is this hard image classification benchmark.
Speaker 3 (41:39):
And I think that showed that.
Speaker 4 (41:40):
The recipe of okay, deep learning, so you take neural
networks plus more computation plus more data, can work very
well for these high dimensional, very difficult problems. And since
then we've seen other ones such as GPT, which you've
just mentioned, and now more recently Alpha folden science, and
I think with people see this big proof of concepts
(42:02):
of one machine learning can do in these large settings
that were very hard to do before, they are naturally
attracted to it. And then we see people either in
industry or in other sciences trying to tackle problems using
machine learning.
Speaker 2 (42:20):
Interesting. Okay, just before we close the goal off, I
wanted to appreciate you taking the time to take this call.
I'm really learning a lot and I hope you're enjoying
this cull too. But before we close the goal off,
could you please recommend me someone to join me or
join us in the next episode of Data Stand Up.
Speaker 3 (42:43):
Yeah.
Speaker 4 (42:43):
So, I think it's always great to chat with people
from various backgrounds. So I think you can maybe reach
out to someone from academia or someone working in other
countries that the guests that you've had so far we're
not working in. So, for example, I know a few
folks doing PhDs and machine learning in Romania and I
might be able to connect you with them if that's
(43:04):
something that you're interested in.
Speaker 2 (43:06):
Yeah, of course. I mean sometimes when you when you
speak with people that are already working in a specific company,
they have those biases of being exposed to a set
of machine learning or dat analytics applications within that industry
and within that company, and there is so much more
(43:27):
to that that we want to learn and we want
to discuss about. So if they are working on new
application and they are expanding this field, we want to know,
we want to know about that, Yeah, of course. And
another question, because you did say that you read a
lot sometimes papers, whatever it is. But give me a book,
(43:52):
a newsletter or something that you sometimes read that you
highly recommend to stay up to the latest in this field,
or maybe because it's a book that's entertaining, and that's
it's really a learning tool for anyone that I know
is trained to have started in something that relates to
(44:14):
a YI or machine learning.
Speaker 4 (44:15):
Yes, as a textbook, I really really like Kevin Murphy's
Machine Learning A Probabilistic Perspective and I always like having
it on my book, so that that's a recommendation, And
he has version two coming up now, so I think
that's that's something to look out for. But I think
for recent work fair and Khosar has an excellent machine
(44:37):
learning blog post at Inference dot VC, and if you're
following the literature closely, I think this is an interesting
blog post to look out for.
Speaker 3 (44:51):
I think it's it's it's a really great resource.
Speaker 2 (44:53):
Okay, fantastic, Thank you so much, Mihaila. I really enjoyed
making this call, and again I appreciate the time that
you took to discuss with this all of all of
these advanstment advancements in in AI. Thank you again, and
I hope you you have a really nice day.
Speaker 3 (45:13):
Thank you so much for having me.
Speaker 2 (45:15):
Thank you Mihailer, have a good one. Bye bye Mu
Popular Podcasts
Stuff You Should Know
If you've ever wanted to know about champagne, satanism, the Stonewall Uprising, chaos theory, LSD, El Nino, true crime and Rosa Parks, then look no further. Josh and Chuck have you covered.
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations. Follow now to get the latest episodes of Dateline NBC completely free, or subscribe to Dateline Premium for ad-free listening and exclusive bonus content: DatelinePremium.com
The Breakfast Club
The World's Most Dangerous Morning Show, The Breakfast Club, With DJ Envy, Jess Hilarious, And Charlamagne Tha God!