episode
stringlengths
45
100
text
stringlengths
1
528
timestamp_link
stringlengths
56
56
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
compressing that data to act efficiently in that
https://karpathy.ai/lexicap/0011-large.html#00:45:48.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
data you yourself appear very often.
https://karpathy.ai/lexicap/0011-large.html#00:45:54.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So it's useful to form compressions of yourself
https://karpathy.ai/lexicap/0011-large.html#00:45:57.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
and it's a really beautiful formulation of what
https://karpathy.ai/lexicap/0011-large.html#00:46:00.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
consciousness is a necessary side effect.
https://karpathy.ai/lexicap/0011-large.html#00:46:03.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
It's actually quite compelling to me.
https://karpathy.ai/lexicap/0011-large.html#00:46:05.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
You've described RNNs, developed LSTMs, long
https://karpathy.ai/lexicap/0011-large.html#00:46:11.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
short term memory networks that are a type of
https://karpathy.ai/lexicap/0011-large.html#00:46:16.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
recurrent neural networks that have gotten a lot
https://karpathy.ai/lexicap/0011-large.html#00:46:20.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
of success recently.
https://karpathy.ai/lexicap/0011-large.html#00:46:23.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So these are networks that model the temporal
https://karpathy.ai/lexicap/0011-large.html#00:46:24.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
aspects in the data, temporal patterns in the
https://karpathy.ai/lexicap/0011-large.html#00:46:27.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
data and you've called them the deepest of the
https://karpathy.ai/lexicap/0011-large.html#00:46:30.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
neural networks.
https://karpathy.ai/lexicap/0011-large.html#00:46:34.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So what do you think is the value of depth in
https://karpathy.ai/lexicap/0011-large.html#00:46:35.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
the models that we use to learn?
https://karpathy.ai/lexicap/0011-large.html#00:46:38.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Since you mentioned the long short term memory
https://karpathy.ai/lexicap/0011-large.html#00:46:41.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
and the LSTM I have to mention the names of the
https://karpathy.ai/lexicap/0011-large.html#00:46:46.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
brilliant students who made that possible.
https://karpathy.ai/lexicap/0011-large.html#00:46:50.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
First of all my first student ever Sepp Hochreiter
https://karpathy.ai/lexicap/0011-large.html#00:46:52.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
who had fundamental insights already in his
https://karpathy.ai/lexicap/0011-large.html#00:46:56.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
diploma thesis.
https://karpathy.ai/lexicap/0011-large.html#00:46:58.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Then Felix Geers who had additional important
https://karpathy.ai/lexicap/0011-large.html#00:46:59.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
contributions.
https://karpathy.ai/lexicap/0011-large.html#00:47:03.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Alex Gray is a guy from Scotland who is mostly
https://karpathy.ai/lexicap/0011-large.html#00:47:04.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
responsible for this CTC algorithm which is now
https://karpathy.ai/lexicap/0011-large.html#00:47:08.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
often used to train the LSTM to do the speech
https://karpathy.ai/lexicap/0011-large.html#00:47:11.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
recognition on all the Google Android phones and
https://karpathy.ai/lexicap/0011-large.html#00:47:15.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
whatever and Siri and so on.
https://karpathy.ai/lexicap/0011-large.html#00:47:18.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So these guys without these guys I would be
https://karpathy.ai/lexicap/0011-large.html#00:47:21.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
nothing.
https://karpathy.ai/lexicap/0011-large.html#00:47:26.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
It's a lot of incredible work.
https://karpathy.ai/lexicap/0011-large.html#00:47:27.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
What is now the depth?
https://karpathy.ai/lexicap/0011-large.html#00:47:29.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
What is the importance of depth?
https://karpathy.ai/lexicap/0011-large.html#00:47:30.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Well most problems in the real world are deep in
https://karpathy.ai/lexicap/0011-large.html#00:47:32.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
the sense that the current input doesn't tell you
https://karpathy.ai/lexicap/0011-large.html#00:47:36.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
all you need to know about the environment.
https://karpathy.ai/lexicap/0011-large.html#00:47:40.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So instead you have to have a memory of what
https://karpathy.ai/lexicap/0011-large.html#00:47:44.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
happened in the past and often important parts of
https://karpathy.ai/lexicap/0011-large.html#00:47:48.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
that memory are dated.
https://karpathy.ai/lexicap/0011-large.html#00:47:51.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
They are pretty old.
https://karpathy.ai/lexicap/0011-large.html#00:47:54.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So when you're doing speech recognition for
https://karpathy.ai/lexicap/0011-large.html#00:47:56.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
example and somebody says 11 then that's about
https://karpathy.ai/lexicap/0011-large.html#00:47:59.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
half a second or something like that which means
https://karpathy.ai/lexicap/0011-large.html#00:48:05.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
it's already 50 time steps.
https://karpathy.ai/lexicap/0011-large.html#00:48:09.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
And another guy or the same guy says 7.
https://karpathy.ai/lexicap/0011-large.html#00:48:11.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So the ending is the same even but now the
https://karpathy.ai/lexicap/0011-large.html#00:48:15.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
system has to see the distinction between 7 and
https://karpathy.ai/lexicap/0011-large.html#00:48:19.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
11 and the only way it can see the difference is
https://karpathy.ai/lexicap/0011-large.html#00:48:22.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
it has to store that 50 steps ago there was an
https://karpathy.ai/lexicap/0011-large.html#00:48:25.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
S or an L, 11 or 7.
https://karpathy.ai/lexicap/0011-large.html#00:48:30.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So there you have already a problem of depth 50
https://karpathy.ai/lexicap/0011-large.html#00:48:34.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
because for each time step you have something
https://karpathy.ai/lexicap/0011-large.html#00:48:37.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
like a virtual layer in the expanded unrolled
https://karpathy.ai/lexicap/0011-large.html#00:48:41.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
version of this recurrent network which is doing
https://karpathy.ai/lexicap/0011-large.html#00:48:44.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
the speech recognition.
https://karpathy.ai/lexicap/0011-large.html#00:48:46.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So these long time lags they translate into
https://karpathy.ai/lexicap/0011-large.html#00:48:48.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
problem depth.
https://karpathy.ai/lexicap/0011-large.html#00:48:51.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
And most problems in this world are such that
https://karpathy.ai/lexicap/0011-large.html#00:48:53.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
you really have to look far back in time to
https://karpathy.ai/lexicap/0011-large.html#00:48:57.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
understand what is the problem and to solve it.
https://karpathy.ai/lexicap/0011-large.html#00:49:01.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
But just like with LSTMs you don't necessarily
https://karpathy.ai/lexicap/0011-large.html#00:49:05.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
need to when you look back in time remember every
https://karpathy.ai/lexicap/0011-large.html#00:49:08.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
aspect you just need to remember the important
https://karpathy.ai/lexicap/0011-large.html#00:49:11.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
aspects.
https://karpathy.ai/lexicap/0011-large.html#00:49:13.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
That's right.
https://karpathy.ai/lexicap/0011-large.html#00:49:14.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
The network has to learn to put the important
https://karpathy.ai/lexicap/0011-large.html#00:49:15.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
stuff into memory and to ignore the unimportant
https://karpathy.ai/lexicap/0011-large.html#00:49:18.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
noise.
https://karpathy.ai/lexicap/0011-large.html#00:49:22.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
But in that sense deeper and deeper is better
https://karpathy.ai/lexicap/0011-large.html#00:49:23.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
or is there a limitation?
https://karpathy.ai/lexicap/0011-large.html#00:49:28.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
I mean LSTM is one of the great examples of
https://karpathy.ai/lexicap/0011-large.html#00:49:30.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
architectures that do something beyond just
https://karpathy.ai/lexicap/0011-large.html#00:49:34.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
deeper and deeper networks.
https://karpathy.ai/lexicap/0011-large.html#00:49:40.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
There's clever mechanisms for filtering data,
https://karpathy.ai/lexicap/0011-large.html#00:49:42.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
for remembering and forgetting.
https://karpathy.ai/lexicap/0011-large.html#00:49:45.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So do you think that kind of thinking is
https://karpathy.ai/lexicap/0011-large.html#00:49:47.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
necessary?
https://karpathy.ai/lexicap/0011-large.html#00:49:50.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
If you think about LSTMs as a leap, a big leap
https://karpathy.ai/lexicap/0011-large.html#00:49:51.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
forward over traditional vanilla RNNs, what do
https://karpathy.ai/lexicap/0011-large.html#00:49:54.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
you think is the next leap within this context?
https://karpathy.ai/lexicap/0011-large.html#00:49:57.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So LSTM is a very clever improvement but LSTM
https://karpathy.ai/lexicap/0011-large.html#00:50:02.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
still don't have the same kind of ability to see
https://karpathy.ai/lexicap/0011-large.html#00:50:06.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
far back in the past as us humans do.
https://karpathy.ai/lexicap/0011-large.html#00:50:10.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
The credit assignment problem across way back
https://karpathy.ai/lexicap/0011-large.html#00:50:14.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
not just 50 time steps or 100 or 1000 but
https://karpathy.ai/lexicap/0011-large.html#00:50:18.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
millions and billions.
https://karpathy.ai/lexicap/0011-large.html#00:50:22.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
It's not clear what are the practical limits of
https://karpathy.ai/lexicap/0011-large.html#00:50:24.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
the LSTM when it comes to looking back.
https://karpathy.ai/lexicap/0011-large.html#00:50:28.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Already in 2006 I think we had examples where
https://karpathy.ai/lexicap/0011-large.html#00:50:31.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
it not only looked back tens of thousands of
https://karpathy.ai/lexicap/0011-large.html#00:50:35.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
steps but really millions of steps.
https://karpathy.ai/lexicap/0011-large.html#00:50:38.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
And Juan Perez Ortiz in my lab I think was the
https://karpathy.ai/lexicap/0011-large.html#00:50:41.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
first author of a paper where we really, was it
https://karpathy.ai/lexicap/0011-large.html#00:50:45.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
2006 or something, had examples where it learned
https://karpathy.ai/lexicap/0011-large.html#00:50:49.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
to look back for more than 10 million steps.
https://karpathy.ai/lexicap/0011-large.html#00:50:53.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So for most problems of speech recognition it's
https://karpathy.ai/lexicap/0011-large.html#00:50:57.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
not necessary to look that far back but there
https://karpathy.ai/lexicap/0011-large.html#00:51:01.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
are examples where it does.
https://karpathy.ai/lexicap/0011-large.html#00:51:05.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Now the looking back thing, that's rather easy
https://karpathy.ai/lexicap/0011-large.html#00:51:07.040