Perhaps [INAUDIBLE] perhaps in a, in a, in
a couple years back, right?
>> Yeah, one and half years back.
>> One and half years and he's one of
the I guess school mentors of, of Markov
logic.
[INAUDIBLE].
>> right, I learned to [INAUDIBLE] but
maybe the
first person who had influence in learning
on that, yeah.
He cares about this interesting area of
integrating.
[COUGH].
>> Logic and statistics.
So welcome, thanks for visiting
>> Thanks Gautham.
Yeah.
So, so this topic is going to be about
Markov logic and I
see some people, who probably would have
at least parts of this talk.
before, but hopefully there'll be
something new, new for you also.
And so here is the brief outline of the
doc.
so, I'll give some motivation for the
language, right?
That Markov logic that will propose.
The background, of some of the languages
which have
been proposed earlier, what is the
relationship within them.
Then I'll formally define the language and
I'll,
describe some inference and learning
[UNKNOWN] to that.
Describe two applications that are well
known, using Markov logic.
And there are plenty of them, that people
have,
all over the world, in fact, using Markov
logic.
And, some future directions, and then
conclude the talk.
Okay.
So, I'll try to give you two different
kinds
of motivations for, for the frame that has
been proposed.
Right?
So this is the first, first motivation.
So ideas of interface layer.
Right, so if you think about any area
[UNKNOWN] then you can think about
applications being developed at one, at
one
friend and also you can think about
infrastructure.
And I'll use some examples and typically
what people have
seen is that, initially when an area sort
of comes.
In, fashion then, you have all these
n-square combinations.
There are n, maybe n different
applications
that you need to worry about and
maybe there are n different so of
infrastructure companies that you need to
worry about.
And for every little developing at the
lowest level you need
to worry about what application, might,
how that application might be affected.
So you have to worry about all this, cross
connections.
And all the time what people realizes
that, maybe there's an interface layer
that.
Separates these two, then application
layer can, really talk to
interface layer and then interface layer
can talk to infrastructure, right?
So this separation really helps instead of
having order n-square
combinations, you will have now, linear
number of combinations, right?
So, just to give an example that will,
make it clearer.
If you think about networking, you have,
applications like www, email and so,
right?
And infrastructure level, contains
protocols, routers, and all such things.
Now, after some time of, you know,
thinking about networking, people thought
that, maybe there is something like
Internet, which really serves as interface
layer.
So what I mean by that, is that, once
you've built to the, sort of the
interface, how.
What the protocols are or what the
how you should, the application should be
communicating.
[COUGH].
>> With the internet then you don't need
to worry about what exactly,
how exactly the Routers, are implemented
or
what is happening at the infrastructure
level.
Let all the application be developed,
almost independently
of what is happening at the lowest level.
Similarly, someone who works at the
infrastructure level doesn't
have to worry about exactly what
applications are been.
Implement it, right?
As long as they are true, to the
interface, to the middle layer, everything
is good, right?
So that is right, yah?
So similarly, when you think about
databases,
you have various things like, ERP, online
transaction
processing systems, CRMs, and many such
applications and,
the infrastructure could be query
optimization, transaction management,.
And various things, you know, where you
really work at the backend.
And what is the interface layer?
Well, people could argue various sort of
things
like maybe relational model or XML but for
large part you could say maybe the
relational
model is something which serves as an
interface layer.
And as, long as you are true to that
schemer, how
we talk with the relational model,
the applications can be developed
independently.
They can be optimized.
They can come up with endless
applications,
without really worrying about the lowest
level.
Similarly people at the lowest level can't
really optimize and
come up with new algorithms, in [UNKNOWN]
or whatever that is.
Without really affecting the top level of,
applications.
Right?
And similarly for A.I. right?
.
So, for A.I. there are a number of
applications.
Robotics, vision, NLP, planning,
multi-agent systems.
And infrastructure of course, corresponds
to knowledge presentation, how do you
do learning, to do inference, right, how
can you make it efficient.
And all those questions, are, are there,
and, over
last you know, 50, 60 years, people have
heard about.
Since that went of A.I. right?
People have talk about what could be the
interface layer.
And then, like, one of the choices that,
people
talked was, maybe first-order logic was
the right choice, right?
Like [UNKNOWN].
Why so?
Because it's very powerful and it can
present entities, relationships, and.
When we talk about high human level
intelligence then
it really is very compact, right, you can
talk
about, groups of people at one, time and
then
you can talk about the properties,
infractions, or all that.
Right.
But then, all the time people realize that
it is a problem with this,
and the problem is that there is no
explicit way to handle [UNKNOWN], right?
There is no way to incorporate the fact
that.
The, rules that you have are noisy.
Right?
And that's, almost always true in the real
world.
So that did not really fly.
Of course, first-order logic is still,
very important languages, but
there are issues in terms of if you have
noisy data
or, if you want to have ex, if you want to
have explicit motion of probability, then,
it doesn't really handle that.
Last 15 or 20 years or maybe, at least,
last 10, 15
years have seen lot of, progress
in something called statistical machine
learning.
Right.
So the methods have become very, very well
founded.
And one set of models which has become
really popular is called graphical models.
Things like Bayesian networks or Markov
networks.
And, people thought that, maybe that gives
you,
an potentially an interface layer for A.I.
right?
And, that also.
Was, quite a bit of success was there, but
again,
now the issue is sort of complimentary to,
sort of logic.
The graphical models in their, in the way
they propose, are proportional.
Right, so you really argue about each
division separately.
There's no.
[COUGH].
>> Way or explicit notion of how to
really handle relations in, into such a
language, right.
And the idea is that.
Can we really combine, the power of these
two
together to potential provide a interface
layer for artificial intelligence?
And I guess next, right, so this is area,
which
has really been for the last 10 years or
so.
So people from statistically and logically
have combined.
Their strengths, and this area called
statistical-relational learning, has come
about, which really combines the power of
logic and probability.
On one side, logic, and probability on the
other side.
And which potentially can provide an
interface layer.
>> So, in the recording [UNKNOWN], that
doesn't mean that you don't have exactly
[UNKNOWN].
I can edit out anything that we don't want
to.
>> Surely, yeah, yeah, so, so feel free to
ask, ask questions.
I guess that is appropriate, yeah.
So I should mention that, at, at least
this part
of the slide has been, borrowed from my
advisor Pedro Domingos.
Right, so, I should, I should mention that
I did, right.
And I guess this, this, is a still, in the
sense this is, I wouldn't really claim
that this is entirely true and, but I will
in this talk I will try to maybe.
Try to, give you some argument that how
Markov Logic which
combines sort of, strength of Markov
networks and infrastructure logic,
infrastructure logic,
could potentially serve as an interface
layer and hopefully when we
look at the inference learning and various
applications that have been developed.
Hopefully we'll get some idea and maybe,
some of you
[INAUDIBLE] is that this can potentially
serve as an interface later.
Of course a long way to go, but it, it's a
good motivation to look at it in this, in
this way.
Right, so, Markov Logic as an interface
layer, potential interface layer,
[INAUDIBLE] you can develop applications,
and then you can develop infrastructure,
almost independently, and as long as you
are familiar with the
semantics of this, of this interface
layer, in this case Markov Logic.
Right.
So this is, this is the first, sort of,
motivation for the talk.
And the second one is, little different.
Coming from a different perspective.
Let's say if you are a pragmatist, you
won't really worry about all
this big bowls of, you know, having an
interface layer for a year, right?
Then, maybe, what you really care about
are, you know, small application.
Little applications and if you can do good
on that.
And so this example is from that.
Right?
So, think about let's say if you have a
social network
analysis, and, you're looking to model
your world, or your domain.
Let's say it is a university, or your
community, where people are smoking.
People have cancer.
People have friendships.
And, you know that smoking leads to,
cancer.
Right, so this is a very good.
[COUGH].
>> Rule of thumb, but we know that this is
not always true.
Right?
People who smoke, they, may or may not
have cancer, but at the same time
we know that, if someone smokes then
they're
more likely to have cancer, and, than not.
Right, so this is a very rough rule of
thumb.
Right?
And again think about, that you don't want
to really go for every little person
your community and your domain saying that
John has, if John smokes, John has cancer.
If Anna smokes, Anna has cancer.
And so on.
Right?
So, you want to, really specify this, very
compactly.
May be some kind of, using some kind of
logic, logical language at the same time
you want
to have this, explicit notion of
uncertainty and we
will, really formally see what that notion
of uncertainty is.
Right?
So similarly you could say that,
friendship leads to similar smoking
habits.
Right, so if I am, friend with lot of
people who smoke, then I am more likely
to.
To smoke that not.
Right?
>> Similarly if I'm friends with a lot of
people who don't
smoke, than I'm more likely to, not be
smoking than, having smoking habits.
Right, so this is again, a very good rule
of thumb, and
this has been verified by social science
that, this in fact is true.
Right?
So how do capture both of these facts?
Right.
That is where bio automatic logic will,
comes in.
Right, so this is, the idea, right?
This is, a goal that we, start with, so
combining logic and probability,
right, so the real world, problem that
are characterised by entities and
relationships, right,
and there is explicit motion of certain
behavior that you want to capture,
so you want to capture both these things,
the relational aspect of the problem.
And the probabilistic aspect of the
problem, right?
Logical languages can handle
relationships, horn clauses, first-order
clauses.
Different logical languages, right, you
could, really choose depending
on your application what might be the
right choice.
Probability can represent uncertainty,
right?
So there is, things like Markov networks,
Bayesian networks.
There are many different statistical,
learning techniques,
but here we'll focus mostly on graphical
models.
Like and then how do we really combine the
two.
Right?
So when I say combine, you don't want,
something that is
heuristic based but you want something
which is really well founded.
Both in, terms of the theory which has
been
developed for logical languages and for
the probabilistic languages.
Right.
So that is the goal.
And if you look at the history of this,
this goes way past in 1980's, right.
There is, something called probabilistic
logic by Nilsson.
And as you see there has been, consistent
progress in
this direction of course I go, in all the
details of
this but just to give you summary,
[UNKNOWN] each language
is different in terms of, the two
components I just listed.
What is the logical language and what is
sort
of the, inference or the statistical part
of it.
Right, for example, if you think about
basic logic programs proposition casting
and loop rate.
Right, in 2001 they combine horn clauses
with Bayesian networks.
Right.
So similar rela, relation Markov networks
[UNKNOWN] the combined
language which is essentially sequel
queries, with Markov networks.
There is a series called BLOG which has,
slightly complicated language
and it combines with the, Bayesian
networks on the, probabilistic side.
And again, there's a lot of literature on
this you can read.
And Markov logic is in fact, is sort of a
latecomer.
It was proposed back in 2006.
There was one short page which came
earlier but the general paper
appeared in 2006 and that combines about a
full first order logic.
For this talk, we'll focus on finite first
order logic.
But there are extensions where you can
really go, to the powerful first order
logic.
And Markov networks, right?
And and I believe there have been, couple
more later, but I won't
really say that Markov logic is probably
the most popular ones among these
and one of the reasons is that it has very
well developed software
system which you can freely download and
use it for your own application.
So that, really has helped to make it
popular and lot
of people as I said in the beginning, not
just at
University of Washington, where it was
proposed but all across the
world people are using it for their own,
application and programs.
Right?
[BLANK_AUDIO]
Okay, so I guess that, ends the
motivation.
I guess it would, maybe a good point
to ask any questions or interrupt at this
point.
And then, if not then I guess I'll
continue with the, with the rest of the
talk.