NIPS 2011 Learning Semantics Workshop: Learning Semantics of Movement

July 25, 2019 posted by

>> Relating to yesterday's presentation. So, there was this kind of question about
formal analysis of language and we did system that produced was answering questions like
this, let me finish. But the basic idea was that you all have the
things, you can–in a [INDISTINCT] form and then, generate [INDISTINCT] of theories [INDISTINCT]
But one of the clear problems was the coverage of [INDISTINCT] mention, so I think that way
we have–and that was also was actually my personal motivation, to move in [INDISTINCT]
doing the rules by then, so. But there were also coordinated problems,
they make problems when [INDISTINCT] as something [INDISTINCT] would be a subject to [INDISTINCT]
so, I very much like Chelsea's–[INDISTINCT] and I have it as–that actually [INDISTINCT]
1991, but the way I [INDISTINCT] so. So, that's not–it's not as sophisticated
so, I have no [INDISTINCT] current–more formal [INDISTINCT] understanding. But one point where I actually changed also
many of the things in [INDISTINCT] that's used in aspects of continuity, which I still
believe that I somehow feel nothing–some of this, it's related to the kind of how contextuality
is being understood in continue–the continuous place referred for example the [INDISTINCT]
theory on conceptual spaces. So, for example, we [INDISTINCT] composition
it is only powerful so I think we–it is interesting thought that like, quantity or grammar was
a nice [INDISTINCT] theory of semantics but, I think it was severely [INDISTINCT] and there
are people also [INDISTINCT] very strong emphasis on [INDISTINCT] And then the [INDISTINCT]
domains has two–like, the example–we have these options here, you know, four options
and then we think that [INDISTINCT] here. But when it's in the complex, we see that
it's either–it's really white and then we categorized it with [INDISTINCT] for its–at
least the program worked up, substance which is there. So, either we get conclusions–is that–especially
when it's really necessary and useful, we will–the coverage and relate it to [INDISTINCT]
it's really usage makes [INDISTINCT] against you make, there was really [INDISTINCT]
>> I'm a regular [INDISTINCT] myself. >> Yes, early also. So–and one of those experiences I've had–I
really became concerned about the strong use of semantic theories on [INDISTINCT] and my
brother [INDISTINCT] about the theory of semantic should be–also there would be fully statistical
[INDISTINCT] in the sense that the group kind of [INDISTINCT] form of representation as
itself is not an appropriate problem for me, but it really requires something else. And one of the conclusions here is that the
supervised learning is often used in order to learn something related values, assuming
a theory of land which even–if I have it here. So, we need to have the context, physical
and social [INDISTINCT] and we don't [INDISTINCT] we don't have regular [INDISTINCT] phrase,
and work phrase and [INDISTINCT] they have to emerge from the–and in that sense, the
experience that [INDISTINCT] large context to be very [INDISTINCT] to learning a lot
of many things, that's why I don't know quite the [INDISTINCT] in the form of [INDISTINCT]
school. So, we are able to learn, so [INDISTINCT]
and I forgot to mention that [INDISTINCT] etcetera, etcetera. And when we take 15-year feet further, the
next thing I mentioned was subjectivity so, we have the program majority about subjectivity. So, I know–I think that [INDISTINCT] we can't
assume that any of us would have the same interpretation basis for anything that we
say or hear. And that’s why we actually have to be enable
to model the conceptual spaces of human beings, we [INDISTINCT] and this is something that
their patient [INDISTINCT] and then go beyond the–when the [INDISTINCT] is close enough
in some financial space, then we are really getting–understanding it all. And then the understanding is not the total
[INDISTINCT] but it's really more of the–more communist base functions. So, next is–step to the Multimodality as
we–so, which takes us towards the movement. So, the–some of our colleagues and–participated
[INDISTINCT] quite successfully be [INDISTINCT] in some of those, then we [INDISTINCT] in
our [INDISTINCT] and go into the details and maybe in the [INDISTINCT] is there and so
[INDISTINCT] research for animation and motion picture so on and so on. But here, I don’t–towards the–from the
topic of–is relation. So many of you may have seen presentation
where this connection that they’ve been taken of, but I re-mentioned it. But actually, for humans and for animals,
a reason for–that we–we have to be able to deal with the movement of which we are
[INDISTINCT] etcetera. And then there's this extreme example of an
animal that transforms itself into a [INDISTINCT] and eating its own small brain. Gross. But my focus is in human movement and what's
interesting is that–is that we won't take it quite from the [INDISTINCT] so, you won't–we
won't thinking about the continuous movement in such a way that the starting point is not
kind of in the categorization but really is around that phenomenon. And one of the things we have such [INDISTINCT]
is that we don’t want to say in that particular kind of movement is working by definition,
or it's running or it's crawling, or whatever but it is and perhaps subjective distribution
over those [INDISTINCT] movements. So, we have [INDISTINCT] of movements, and
that’s–and actually [INDISTINCT] space and then we have [INDISTINCT] using sheets
or words and expressions or the–and none of those usage of words is taken as a grounds
rule, but it's more like the distribution that we want to learn about in the end. That's the [INDISTINCT] interrupts the movement
to what [INDISTINCT] spaces are hand movement and to use petronets to formalize the use
of manual way of determining some of those underlying concepts. And we actually want to give more answers
of what's [INDISTINCT] let's supervise in the sense [INDISTINCT] in the theory formation. But anyway, [INDISTINCT] walking is [INDISTINCT]
as multiple dimensions. And then there are people who have like, [INDISTINCT]
who has tried to cut around, you know, [INDISTINCT] so this stuff works like [INDISTINCT] And
the [INDISTINCT] is to link with language and movement in such a way that he–for example
use language to generate animation and movement in such a way, well then, we can–for example
prompt the expressions in different languages, using movement as a [INDISTINCT] so that single
space of [INDISTINCT] movements would be basis for much [INDISTINCT] So now, I am [INDISTINCT]
>> Thank you. So I'm going to talk about some of our ongoing
work on learning relations. So first, many works that we use don’t describe
object category as properties of [INDISTINCT] reading our [INDISTINCT] or whatever. But rather a combination of objects and–so,
let me explain [INDISTINCT] conceptual relations to [INDISTINCT] for example, this is my one
fist and the other fist follows it, then what can we really say about [INDISTINCT] well,
we need a–we know that–if A follows B, then we know that–if we know the location of B
and then A is also nearby with a very [INDISTINCT] And we can also know that we don’t need
to know very much about A to be able to infer this. And another example is container so, if we
have some kind of container like a bag for example, we can take any object and as long
as the containment is relation, who else–we can move the bag wherever we want and when
we open the bag, the thing which is inside is obviously there. We don' know this but–yeah. So, we have tried to formalize this in such
a way that–[INDISTINCT] definition of relation is that, a relation is an end place for you
to get that it's either true or false, or some arguments, so. We have A and B, and relation R as the rules
are [INDISTINCT] And they have said that rules that allow reasoning so, if you say that you
know that A follows B, then you might follow that–if we know that [INDISTINCT] B then,
it follows that 0 and Y and A as well. And this is quite difficult to apply in a
perceptional [INDISTINCT] setting because this is probably going to be [INDISTINCT]
so we have turned these definitions slightly on its head and say that the important part
is actually the prediction. So, if we can predict something from–we know–let's
take the form the beginning, so when we know that the relation falls between X and Y then,
we can predict something about X from Y that we can't predict from X alone. And we do this exactly symmetrically so there
it is, the definition as symmetrically, but we implement this exactly symmetrically. So, we have a learning average for a movement
data where we have X and Y which represent different objects. And then we learn relations using a [INDISTINCT]
that observes the changing correlation structure between the variables. So, at each time an interrelation, there is
a correlation seeing on [INDISTINCT] that deals as how active is the relation for a
particular time. And this produces a very fast [INDISTINCT]
but so far we have only applied [INDISTINCT] motion and I'll show you an example. So, we have these green agents and they are–they
are roaming around near in this space where they live and some of them are hungry and
are chasing these cakes. So, you can see all the cakes are moving. And that does happen very often in the real
world though. Here in the simulation, these cakes are maybe
going bad or something. So, you see that this guy is going here towards
the corner and there's no cake here so it's probably not hungry. But this guy on the other hand, you see he
started out here and then he [INDISTINCT] and he's getting [INDISTINCT] his probably
target in that cake. So, tries figure out when–which spares of
agents are targeting a cake at that particular point in time. And here is–a results of the greenish that
[INDISTINCT] so, it's one when the relation is known to be active and then it's zero otherwise. And the blue one is the estimated RPC, you
know. So, we see that the [INDISTINCT] to zero are
quite noisy but, but maybe it's active only [INDISTINCT] relation is actually [INDISTINCT]
So, the point about learning relations is really that it allows us to generalize for
completely [INDISTINCT] objects there, so that–if I know that something is following
something else, I don’t need to know anything about the object categories really, except
that maybe there are physical objects. If there's an alien from outer space and it's
following me I can't [INDISTINCT] pretty well. And then related to much that I has been said
and this works on the [INDISTINCT] saw the Metaphoric relation generalization which something
that is known to happening in which you apply one concept from one domain to an [INDISTINCT]
different domain that you can describe [INDISTINCT] movement, like butterflies or something like
that. Maybe that is some kind of automatic discovery
of compatible that formality has to learn. So, that would be A in this relation, maybe
one-step further. >> Maybe you can show that previous slide
in photos [INDISTINCT] >> Sure.

No Comments

Leave a Comment

Your email address will not be published. Required fields are marked *