title
stringlengths 12
112
| published
stringlengths 19
23
| url
stringlengths 28
28
| video_id
stringlengths 11
11
| channel_id
stringclasses 5
values | id
stringlengths 16
31
| text
stringlengths 0
596
| start
float64 0
37.8k
| end
float64 2.18
37.8k
|
|---|---|---|---|---|---|---|---|---|
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t315.72
|
a system that when we input one image, it can give us one parse tree. But when we input
| 315.72 | 326.28 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t321.36
|
another image, it can give us some kind of other parse tree, maybe now there are two
| 321.36 | 334.96 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t326.28000000000003
|
objects in the image. And this one has one descendant only, which in turn has two descendants,
| 326.28 | 341.68 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t334.96000000000004
|
and so on, you see the point, the tree structure needs to be different each time. This in part
| 334.96 | 348.2 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t341.68
|
was addressed by Hinton's capsule networks. So in the capsule networks, Hinton's idea
| 341.68 | 353.08 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t348.2
|
was sort of, okay, I'm going to have these capsules here in different layers. And I'm
| 348.2 | 359.72 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t353.08
|
going to have kind of lots of capsules and these layers, lots of capsules in these layers.
| 353.08 | 366.08 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t359.72
|
And I'm going over capsules, because it's kind of important here. So Hinton's idea with
| 359.72 | 373.7 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t366.08
|
capsules was that the first layer of capsules would sort of recognize the smallest parts.
| 366.08 | 379.68 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t373.7
|
So this would be kind of the wheel capsule. And this would be sort of the window capsule,
| 373.7 | 385.16 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t379.68
|
and so on. So there would be a single capsule for every part that could possibly be in an
| 379.68 | 390.88 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t385.15999999999997
|
image, right? You already see the limitations. Because if you want to recognize the whole
| 385.16 | 398.16 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t390.88
|
world, you need many capsules. But nevertheless, this was the idea. So a capsule would be active
| 390.88 | 404.92 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t398.15999999999997
|
if there was the given object in the image. And then the next thing here, this would be
| 398.16 | 412.4 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t404.92
|
kind of the the motor capsule. So the motor, motor capsule, and this would be the cabin
| 404.92 | 420.26 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t412.4
|
capsule, and so on. So the window would activate the cabin capsule, but the door capsule would
| 412.4 | 426.02 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t420.26
|
also activate the cabin capsule, and so on. And the wheel would maybe activate, it would
| 420.26 | 432.08 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t426.02
|
maybe activate, I don't know, the wheel should probably be here as well, wheel at this level,
| 426.02 | 438.08 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t432.08
|
would activate that. And then all of these things here would activate the car capsule,
| 432.08 | 446.32 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t438.08
|
sorry. So you can see that this parse tree here is generated dynamically, right? These
| 438.08 | 451.36 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t446.32
|
connections, this routing and capsules is generated every time different. So in the
| 446.32 | 455.44 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t451.36
|
next image, there could be a different object, different capsules are activated, different
| 451.36 | 459.68 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t455.44
|
things are routed together, the parse tree is different. However, you need these many,
| 455.44 | 466.96 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t459.68
|
many capsules for that every one capsule per possible part in the image. And that was just
| 459.68 | 473.88 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t466.96
|
infeasible. And also the routing was very cumbersome in these capsules. So here we go
| 466.96 | 482.92 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t473.88
|
with a new approach. And this new approach is what Hinton describes as the glom architecture
| 473.88 | 489.76 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t482.92
|
is composed of a large number of columns, which all use exactly the same weight. Each
| 482.92 | 496 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t489.76
|
column is a stack of spatially local auto encoders that learn multiple levels of representation
| 489.76 | 503.12 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t496.0
|
for what is happening in a small image patch. Okay, so we're going to build up some kind
| 496 | 508.24 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t503.12
|
of imagination here. At the at the bottom level, we have our image. So our image is
| 503.12 | 515.04 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t508.24
|
going to be lying flat on the ground, maybe you can see like this. And it is going to
| 508.24 | 521.22 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t515.04
|
be divided into pixels or small patches, whatever you want. But these are would be called locations.
| 515.04 | 530.74 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t521.22
|
So it would be divided like this into different locations. I am not good at perspective drawing.
| 521.22 | 536.66 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t530.74
|
In any case, above each location, there would be one of these columns. And these columns,
| 530.74 | 546.44 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t536.66
|
I can draw one here, these columns would sort of stack up like this. And these columns would
| 536.66 | 550.68 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t546.44
|
be divided into multiple levels. So there would be a bottom level, which would be this
| 546.44 | 557.28 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t550.6800000000001
|
there will be a middle level, higher level, and so on. Hinton suggests about five levels
| 550.68 | 566.48 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t557.28
|
should probably do. And every single level of this column tries to represent the location
| 557.28 | 573.84 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t566.48
|
at the image, right, this location down here in a different resolution. So the very bottom
| 566.48 | 580.6 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t573.8399999999999
|
level might be aware that there is a part of a wheel, like let's say this is actually
| 573.84 | 593.56 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t580.6
|
let's say this is a cat. I so here, there's probably Yep, yep. Okay, so you can see, there
| 580.6 | 604.68 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t593.5600000000001
|
is there is an ear or a part of an ear that stays there's a part of an ear in this location.
| 593.56 | 610.58 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t604.6800000000001
|
So the very bottom thing would probably represent something like the very structure of the first
| 604.68 | 615.92 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t610.58
|
layer. So the bottom thing would represent what's going on at you know, the micro level,
| 610.58 | 621.96 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t615.9200000000001
|
really the location level, the next layer would represent what's going on at this location
| 615.92 | 626.32 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t621.96
|
in a kind of a broader sense. So that might recognize that that that's an that's actually
| 621.96 | 632.64 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t626.32
|
part of an ear, right? So it goes beyond the location. If you think convolutional neural
| 626.32 | 637.52 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t632.64
|
networks, you're in the right ballpark. But we're going to implement this differently.
| 632.64 | 648.68 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t637.52
|
The next layer will recognize well, this location is part of a of a cat of a cat's head. And
| 637.52 | 654.64 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t648.68
|
then the next location will recognize well, this thing is part of a cat. So at this location,
| 648.68 | 660.92 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t654.64
|
there's a cat that there there is a cat at other places. But at this location, there
| 654.64 | 667.48 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t660.92
|
is a cat, and so on. So maybe we don't have more and this look at this particular image.
| 660.92 | 676.02 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t667.4799999999999
|
But if you consider a different column, like this, this column right here. And you look
| 667.48 | 681.72 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t676.02
|
at what's going on in that column, you'll see similar. So in the top layer, let's just
| 676.02 | 687.54 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t681.7199999999999
|
consider the cat the top layer in the top layer, it might say, well, there's a cat too.
| 681.72 | 697.68 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t687.54
|
But it's also part of it's part of a cat's neck, neck. And then here, it's maybe there's
| 687.54 | 706.04 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t697.68
|
a bunch of well, I don't know a chin. And there is also a fine first structure of the
| 697.68 | 713.1 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t706.04
|
chin. So you get the idea every column will build up these repre these representations.
| 706.04 | 718.08 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t713.1
|
And these are vectors. So these are embedding vectors. So at the bottom location, you'd
| 713.1 | 724.38 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t718.08
|
have the fur vector, and then this vector is the ear, whereas here over here, the chin
| 718.08 | 730.4 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t724.38
|
would be very different, it will be a different vector at the same layer. So the only thing
| 724.38 | 736.64 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t730.4
|
that agrees here is the cat vector, the cat vector in this top layer would agree between
| 730.4 | 742 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t736.64
|
both of these columns. I hope you get the idea, you have a column above each of the
| 736.64 | 748.38 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t742.0
|
locations, every single layer in the column represents that particular location, but at
| 742 | 755.04 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t748.38
|
a different level of abstraction and a different level of I don't want to say resolution, but
| 748.38 | 761.16 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t755.04
|
it, it would consider more and more of its neighbors. The question is, how does it consider
| 755.04 | 768.4 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t761.16
|
its neighbors? And how do you learn these things, right? So how do you learn these different
| 761.16 | 775.66 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t768.4
|
abstractions? And that's where these columns, they communicate with each other. So Hinton
| 768.4 | 783.78 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t775.66
|
imagines that this is a process over time, where the columns iteratively communicate
| 775.66 | 789.48 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t783.78
|
to each other. And within the column, the layers communicate to each other. And this
| 783.78 | 797.12 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t789.48
|
is one of these first diagrams right here. So this is one single column over time, okay,
| 789.48 | 804.98 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t797.12
|
this is this would be the this would be the fur at the ear, this would be the cat's ear,
| 797.12 | 816.4 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t804.98
|
and this would be cat. Okay, so the information that so the embeddings are updated by sending
| 804.98 | 822.64 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t816.4
|
information around every single embedding, which means that every single vector at every
| 816.4 | 830.64 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t822.64
|
single layer of every single column is updated by simply averaging four things. So we have
| 822.64 | 840.36 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t830.64
|
the embedding at layer l, at time step t plus one is going to be sorry, at layer l location
| 830.64 | 848.96 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t840.36
|
x is going to be a sum between the four parts, the four following parts, it's going to be
| 840.36 | 856.36 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t848.96
|
the embedding at the last time step, right. So this is sort of a recurrent neural network.
| 848.96 | 865.62 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t856.36
|
We the new embedding is the old embedding, plus, it's going to be a function at a top
| 856.36 | 873.4 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t865.62
|
down, this what Hinton calls top down function of the embedding at the same location in the
| 865.62 | 882.76 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t873.4
|
previous time step at one layer above, so l plus one, it is also going to be receiving
| 873.4 | 890.36 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t882.76
|
information from the upwards, I think bottom up, because the bottom up embedding of layer
| 882.76 | 897.4 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t890.36
|
l minus one at the same location at time step t. Alright, so this way, that's what you can
| 890.36 | 906.8 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t897.4
|
see right here, the green arrows are each level each layer simply passes information
| 897.4 | 914.84 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t906.8
|
to the next time step, this is if any if nothing else happens, you just keep your embedding,
| 906.8 | 923.6 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t914.84
|
then each embedding also sends itself through a neural network, one layer above itself,
| 914.84 | 930.08 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t923.6
|
that's the blue arrows. So the blue arrows here are these and you every everything is
| 923.6 | 934.84 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t930.08
|
a neural network here every arrow except the green ones, but the green ones could be too.
| 930.08 | 941.86 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t934.84
|
So every arrow is a neural network. So this is a neural network sending information above.
| 934.84 | 947.88 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t941.86
|
And this is intuitive, right? So the ear embedding would sort of send information about itself
| 941.86 | 954.96 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t947.88
|
like saying like, Hey, I'm a cat ear sends it above and it goes through a neural network
| 947.88 | 960.68 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t954.96
|
because it needs to be transformed. The neural network has to learn well, if it's a cat ear
| 954.96 | 969.16 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t960.68
|
at that level, it might be a cat at the top level. And lastly, every single layer sends
| 960.68 | 974.98 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t969.16
|
information down and that is the red arrows right here. They're also neural networks.
| 969.16 | 982.16 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t974.98
|
So the cat ear says, Well, I'm a cat ear. So downstream of myself, there might be, you
| 974.98 | 988.4 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t982.16
|
know, some first structure, right? So all of these embeddings, they try to predict each
| 982.16 | 994.96 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t988.4
|
other, they try to predict the neighbors of themselves. And Hinton's idea is that by aggregating
| 988.4 | 1,003.3 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t994.96
|
over time, they will sort of reach a consensus of what is in these columns. There are a few
| 994.96 | 1,008.54 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t1003.3
|
things missing right here. The one thing that's missing and hint and pointed this out that
| 1,003.3 | 1,015.28 |
GLOM: How to represent part-whole hierarchies in a neural network (Geoff Hinton's Paper Explained)
|
2021-02-27 15:47:03
|
https://youtu.be/cllFzkvrYmE
|
cllFzkvrYmE
|
UCZHmQk67mSJgfCCTn7xBfew
|
cllFzkvrYmE-t1008.54
|
all of these different columns that we've drawn, they use the same weights. Okay, so,
| 1,008.54 | 1,020.06 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.