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-t1015.28
|
and he discusses this at the end of the paper, it's not really biologically plausible, but
| 1,015.28 | 1,026.84 |
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-t1020.06
|
there's an ensemble effect, we won't go into that. But all these decent so the blue arrows
| 1,020.06 | 1,033.68 |
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-t1026.84
|
are always the same for each time step. But not necessarily the same between different
| 1,026.84 | 1,040.16 |
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-t1033.6799999999998
|
layers. So that might be this f might be different from this f down here. However, the function
| 1,033.68 | 1,045.96 |
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-t1040.1599999999999
|
passing information from from layer l to layer l plus one is the same in every single column
| 1,040.16 | 1,051 |
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-t1045.9599999999998
|
across the image, it's a bit like a convolutional network in terms of weight sharing. So you
| 1,045.96 | 1,057.26 |
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-t1051.0
|
can imagine it as one by one convolutional network in that sense. But except the information
| 1,051 | 1,063.64 |
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-t1057.26
|
does not only go up the layers, it also goes down the layers over time. As I said, this
| 1,057.26 | 1,070.76 |
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-t1063.64
|
is an iterative procedure, goes up, down, and laterally. The second thing is now that
| 1,063.64 | 1,078.02 |
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-t1070.76
|
you ask, Oh, well, if every single column has the same weights, wouldn't that simply
| 1,070.76 | 1,085.4 |
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-t1078.02
|
sort of how how can you localize any information? And the answer is that you have a side input,
| 1,078.02 | 1,090.6 |
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-t1085.4
|
like in a neural field, you have a side input annotating each location, basically a positional
| 1,085.4 | 1,098.12 |
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-t1090.6
|
encoding, honestly. So in in addition to what the image patch looks like, you also get your
| 1,090.6 | 1,104.04 |
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-t1098.12
|
kind of either your x y coordinates, or you could also get your relative coordinates to
| 1,098.12 | 1,112.36 |
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-t1104.04
|
some other coordinate frame in there. And so the network knows where it is. And that's
| 1,104.04 | 1,119.96 |
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-t1112.36
|
going to be important, because what Hinton wants to build are these islands. So the imagination
| 1,112.36 | 1,127.48 |
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-t1119.96
|
of Hinton is that this is going to be somewhere in between like after time step 10. And you
| 1,119.96 | 1,134.72 |
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-t1127.48
|
want to run it for 100. And he imagines that there will what will emerge are these sort
| 1,127.48 | 1,142.5 |
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-t1134.72
|
of islands. So imagine the image is now a 1d vector down here. Or you can imagine these
| 1,134.72 | 1,149.72 |
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-t1142.5
|
columns in 2d, whatever fits, you know, whatever fits your brain better. But imagine the images,
| 1,142.5 | 1,156.12 |
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-t1149.72
|
the image is simply a 1d line right here. He imagines that the bottom vectors, they
| 1,149.72 | 1,162.6 |
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-t1156.12
|
will just, you know, happily kind of be describing whatever that is at the very bottom level.
| 1,156.12 | 1,169.08 |
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-t1162.6
|
But then at the next level, once it goes to sort of higher resolution or lower resolution,
| 1,162.6 | 1,177.52 |
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-t1169.08
|
higher abstraction, there will be there must necessarily be vectors that are the same.
| 1,169.08 | 1,182.32 |
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-t1177.52
|
If the system works and look at these two vectors and look at these two vectors, they
| 1,177.52 | 1,187.68 |
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-t1182.32
|
are the same because they now describe objects that are larger than one location, right,
| 1,182.32 | 1,194.36 |
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-t1187.6799999999998
|
the cat's head is larger than simply one location. Therefore, at the layer that represents the
| 1,187.68 | 1,201.52 |
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-t1194.36
|
cat's head, you expect because these are all all neuron, all the up and down functions
| 1,194.36 | 1,207.86 |
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-t1201.52
|
in the same layer have the same weight, you expect that the embedding of a cat's head
| 1,201.52 | 1,216.02 |
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-t1207.86
|
is the same in in the different columns. This is if the system works, this must be the case.
| 1,207.86 | 1,221.12 |
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-t1216.02
|
And then as you go up, you expect more and more of these what what hint calls islands
| 1,216.02 | 1,230.8 |
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-t1221.12
|
to emerge, right? So they they agree. And the idea the idea between all of this message
| 1,221.12 | 1,238.68 |
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-t1230.8
|
passing is that over time, all of these things kind of reinforce each other. So we looked
| 1,230.8 | 1,246.2 |
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-t1238.68
|
at a column before, and we maybe said, Okay, so this vector down here, it gets information
| 1,238.68 | 1,252.96 |
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-t1246.2
|
from the top, saying, Hey, you know, there's a cat here. So you might be like a cat ear
| 1,246.2 | 1,257.84 |
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-t1252.96
|
or a cat eye or something like this. And then it gets information from the bottom saying,
| 1,252.96 | 1,262.76 |
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-t1257.84
|
well, there's a bit of as you know, fur here, and there's some cartilage showing and so
| 1,257.84 | 1,269.84 |
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-t1262.76
|
on. And it has already sort of figured out that it might be an ear. And these informations,
| 1,262.76 | 1,274.04 |
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-t1269.84
|
they own they reinforce itself now, like they'd be like, okay, you know, you're saying I'm
| 1,269.84 | 1,278.48 |
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-t1274.04
|
part of a head and you're saying there's a bit of fur and cartilage. And I already kind
| 1,274.04 | 1,284.76 |
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-t1278.48
|
of noticed that I'm a bit like an ear. So I'm probably more an ear. So the idea is that
| 1,278.48 | 1,291.2 |
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-t1284.76
|
over time, you have this consensus algorithm, there's one thing missing. And that is, how
| 1,284.76 | 1,296.36 |
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-t1291.2
|
do the different columns communicate with each other. So I said there are different
| 1,291.2 | 1,304.44 |
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-t1296.36
|
parts, there is one missing. And that one missing is going to be, I'm just going to
| 1,296.36 | 1,314.3 |
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-t1304.44
|
call it whatever a and a is going to be an attention mechanism across all the other columns
| 1,304.44 | 1,320.18 |
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-t1314.3
|
at the same layer. So if we look here, this cell receives information from above from
| 1,314.3 | 1,327.48 |
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-t1320.18
|
below from itself. And also, in an attention mechanism way, it's going to receive information
| 1,320.18 | 1,335.72 |
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-t1327.48
|
from all of the different, all of the different embeddings at the same layer. You can see
| 1,327.48 | 1,346.84 |
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-t1335.72
|
that, you know, it puts in everything we got in here. Now, the attention, he says, is easier.
| 1,335.72 | 1,354.16 |
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-t1346.84
|
And so these are the four parts right here. At each discrete time, and in each column
| 1,346.84 | 1,359.8 |
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-t1354.16
|
separately, the embedding at a level is updated to be the weighted average of four contributions.
| 1,354.16 | 1,365.28 |
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-t1359.8
|
The prediction produced by the bottom up neural net acting on the embedding at the level below
| 1,359.8 | 1,372.58 |
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-t1365.28
|
at the previous time, the prediction produced by the top down neural net acting on the embedding
| 1,365.28 | 1,378.84 |
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-t1372.58
|
at the level above at the previous time, the embedding vector at the previous time step,
| 1,372.58 | 1,384.56 |
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-t1378.84
|
these three we got, and then the attention weighted average of the embeddings at the
| 1,378.84 | 1,394.72 |
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-t1384.56
|
same level, right at the same level in nearby columns at the previous time. So nearby heat,
| 1,384.56 | 1,401.8 |
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-t1394.72
|
sorry, he later backpedals a bit, I think on nearby and what nearby exactly means. And
| 1,394.72 | 1,407.6 |
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-t1401.8
|
he some parts. So this, this is idea, I think this is still up for debate. And this is,
| 1,401.8 | 1,413.96 |
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-t1407.6000000000001
|
I think, where I can help. But what he wants to do is he wants to aggregate, he wants to
| 1,407.6 | 1,421.44 |
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-t1413.96
|
attention aggregate, and he wants to simplify attention. So instead, what we usually have
| 1,413.96 | 1,429.04 |
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-t1421.44
|
is we're going to produce queries, and keys and values, queries, keys and values. And
| 1,421.44 | 1,436.24 |
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-t1429.04
|
they're all going to be different functions of our input. And then we're going to do query
| 1,429.04 | 1,443.1 |
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-t1436.24
|
times key transposed softmax of that times value. And that is going to be our attention
| 1,436.24 | 1,448.12 |
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-t1443.1000000000001
|
mechanism that allows you know, arbitrary information to be routed around and so on.
| 1,443.1 | 1,454.92 |
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-t1448.12
|
Attention says, Nope, what I want is simply that all the queries, the keys and the values,
| 1,448.12 | 1,464.76 |
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-t1454.9199999999998
|
they're all just equal to the embeddings themselves. So the attention mechanism would work out
| 1,454.92 | 1,477.52 |
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-t1464.76
|
to be the softmax of x times x transposed times x. And what that does is if you yourself
| 1,464.76 | 1,485.28 |
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-t1477.52
|
are the query, and every vector also itself is the key, what do you attend to, you attend
| 1,477.52 | 1,492.36 |
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-t1485.28
|
to vectors that are very similar to yourself. And you can see that in Hinton's diagram,
| 1,485.28 | 1,497.88 |
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-t1492.36
|
the one we circled dark blue, what would it attend to? Well, it would probably attend
| 1,492.36 | 1,504.76 |
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-t1497.8799999999999
|
to its left hand neighbor, the one you can see circled, I'm going to circle it. This
| 1,497.88 | 1,510.6 |
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-t1504.76
|
one, it will probably attend a lot to this one, it might not attend so much. And the
| 1,504.76 | 1,516.72 |
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-t1510.6
|
ones over here, it might not attend at all. What does this give us, especially since the
| 1,510.6 | 1,524.36 |
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-t1516.72
|
values are also these vectors, this is a consensus algorithm, it is not meant as a way to pass
| 1,516.72 | 1,529.56 |
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-t1524.36
|
information around, it is not meant like in a transformer as a way to do computation,
| 1,524.36 | 1,535.76 |
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-t1529.56
|
because we have no trainable weights in this process. It is simply meant as a consensus
| 1,529.56 | 1,543.88 |
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-t1535.76
|
algorithm. So it imagines that by doing this, by sort of attending to things that are similar
| 1,535.76 | 1,549.96 |
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-t1543.8799999999999
|
to you, and then integrating their values, there will be these islands forming. And that's
| 1,543.88 | 1,555 |
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-t1549.96
|
what you see right here, you can imagine if two vectors are already close at the same
| 1,549.96 | 1,561.64 |
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-t1555.0
|
year, this mechanism will make them even closer. So this is a sort of a clustering algorithm.
| 1,555 | 1,570.6 |
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-t1561.64
|
And so that my question is, that these drawings, you look at them, they are very specifically
| 1,561.64 | 1,577.74 |
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-t1570.6
|
constructed, they are constructed such that a parse tree is emerging. So when you look
| 1,570.6 | 1,585.28 |
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-t1577.74
|
at this, you have a clear sense I can probably, I can probably move all of that crap out of
| 1,577.74 | 1,593.84 |
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-t1585.28
|
the way. You can see the parse tree, right? Because the black thing is going to be the
| 1,585.28 | 1,597.8 |
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-t1593.84
|
top node right here. Let's leave away the scene level embedding for now, the black thing
| 1,593.84 | 1,605.16 |
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-t1597.8
|
is going to be the top node. And then it has two child nodes, this one, and this one. And
| 1,597.8 | 1,610.32 |
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-t1605.16
|
then it has four, every one of those has two child nodes. But it's not it doesn't have
| 1,605.16 | 1,614.96 |
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-t1610.3200000000002
|
to be in this case. So this dynamically and every one of them, you know, the black ones
| 1,610.32 | 1,620.96 |
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-t1614.96
|
are individual. This is dynamically constructing a parse tree, right? The parse tree here
| 1,614.96 | 1,632.52 |
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-t1620.96
|
is something like this. And then the da da da. So this is pretty cool. But it is also
| 1,620.96 | 1,638.4 |
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-t1632.52
|
drawn deliberately such that a core problem does not arise. And the core problem would
| 1,632.52 | 1,646.44 |
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-t1638.4
|
be something like, well, what if this vector here was actually also pointing like this?
| 1,638.4 | 1,652.96 |
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-t1646.44
|
Okay, so it is not in it is not in the same, it is not in the same area of the parse tree,
| 1,646.44 | 1,660.84 |
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-t1652.96
|
right? If you go down the parse tree, it is actually here. Now, if we do what Hinton says,
| 1,652.96 | 1,669.32 |
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-t1660.84
|
and if for this vector here, we do this aggregation via attention on the same layer, what we will
| 1,660.84 | 1,676.8 |
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-t1669.32
|
attend to is this vector over here. Now, this is probably not meant to be because this vector
| 1,669.32 | 1,682.4 |
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-t1676.8
|
over here, it can represent the same thing. But you can see it's not in the in the same
| 1,676.8 | 1,690.08 |
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-t1682.3999999999999
|
path of the parse tree. And he mentioned this a little bit throughout, but not necessarily
| 1,682.4 | 1,697.28 |
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-t1690.08
|
clear. And the drawing makes it seem like there's no problem. But I hope you can see
| 1,690.08 | 1,702.82 |
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-t1697.28
|
how this is a problem. The attention would pull in information from over here. However,
| 1,697.28 | 1,707.42 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.