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
at least all the solvable problems.
https://karpathy.ai/lexicap/0011-large.html#00:02:22.680
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So how do you think, what is the mechanism
https://karpathy.ai/lexicap/0011-large.html#00:02:26.080
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
for that kind of general solver look like?
https://karpathy.ai/lexicap/0011-large.html#00:02:28.120
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Obviously we don't quite yet have one
https://karpathy.ai/lexicap/0011-large.html#00:02:31.640
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
or know how to build one but we have ideas
https://karpathy.ai/lexicap/0011-large.html#00:02:34.840
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
and you have had throughout your career
https://karpathy.ai/lexicap/0011-large.html#00:02:37.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
several ideas about it.
https://karpathy.ai/lexicap/0011-large.html#00:02:39.120
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So how do you think about that mechanism?
https://karpathy.ai/lexicap/0011-large.html#00:02:40.800
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So in the 80s, I thought about how to build this machine
https://karpathy.ai/lexicap/0011-large.html#00:02:43.640
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
that learns to solve all these problems
https://karpathy.ai/lexicap/0011-large.html#00:02:48.640
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
that I cannot solve myself.
https://karpathy.ai/lexicap/0011-large.html#00:02:51.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
And I thought it is clear it has to be a machine
https://karpathy.ai/lexicap/0011-large.html#00:02:54.120
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
that not only learns to solve this problem here
https://karpathy.ai/lexicap/0011-large.html#00:02:57.160
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
and this problem here but it also has to learn
https://karpathy.ai/lexicap/0011-large.html#00:03:00.880
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
to improve the learning algorithm itself.
https://karpathy.ai/lexicap/0011-large.html#00:03:04.160
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So it has to have the learning algorithm
https://karpathy.ai/lexicap/0011-large.html#00:03:09.360
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
in a representation that allows it to inspect it
https://karpathy.ai/lexicap/0011-large.html#00:03:12.480
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
and modify it such that it can come up
https://karpathy.ai/lexicap/0011-large.html#00:03:15.720
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
with a better learning algorithm.
https://karpathy.ai/lexicap/0011-large.html#00:03:19.240
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So I call that meta learning, learning to learn
https://karpathy.ai/lexicap/0011-large.html#00:03:21.080
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
and recursive self improvement
https://karpathy.ai/lexicap/0011-large.html#00:03:24.600
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
that is really the pinnacle of that
https://karpathy.ai/lexicap/0011-large.html#00:03:26.760
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
where you then not only learn how to improve
https://karpathy.ai/lexicap/0011-large.html#00:03:28.760
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
on that problem and on that
https://karpathy.ai/lexicap/0011-large.html#00:03:34.800
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
but you also improve the way the machine improves
https://karpathy.ai/lexicap/0011-large.html#00:03:36.440
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
and you also improve the way it improves
https://karpathy.ai/lexicap/0011-large.html#00:03:40.000
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
the way it improves itself.
https://karpathy.ai/lexicap/0011-large.html#00:03:42.160
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
And that was my 1987 diploma thesis
https://karpathy.ai/lexicap/0011-large.html#00:03:44.600
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
which was all about that higher education
https://karpathy.ai/lexicap/0011-large.html#00:03:47.440
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
hierarchy of meta learners that have no computational limits
https://karpathy.ai/lexicap/0011-large.html#00:03:50.920
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
except for the well known limits that Gödel identified
https://karpathy.ai/lexicap/0011-large.html#00:03:57.240
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
in 1931 and for the limits of physics.
https://karpathy.ai/lexicap/0011-large.html#00:04:01.640
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
In the recent years, meta learning has gained popularity
https://karpathy.ai/lexicap/0011-large.html#00:04:06.480
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
in a specific kind of form.
https://karpathy.ai/lexicap/0011-large.html#00:04:10.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
You've talked about how that's not really meta learning
https://karpathy.ai/lexicap/0011-large.html#00:04:12.760
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
with neural networks, that's more basic transfer learning.
https://karpathy.ai/lexicap/0011-large.html#00:04:16.000
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Can you talk about the difference
https://karpathy.ai/lexicap/0011-large.html#00:04:21.480
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
between the big general meta learning
https://karpathy.ai/lexicap/0011-large.html#00:04:22.720
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
and a more narrow sense of meta learning
https://karpathy.ai/lexicap/0011-large.html#00:04:25.460
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
the way it's used today, the way it's talked about today?
https://karpathy.ai/lexicap/0011-large.html#00:04:27.960
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Let's take the example of a deep neural network
https://karpathy.ai/lexicap/0011-large.html#00:04:30.880
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
that has learned to classify images
https://karpathy.ai/lexicap/0011-large.html#00:04:33.440
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
and maybe you have trained that network
https://karpathy.ai/lexicap/0011-large.html#00:04:37.240
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
on 100 different databases of images.
https://karpathy.ai/lexicap/0011-large.html#00:04:40.060
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
And now a new database comes along
https://karpathy.ai/lexicap/0011-large.html#00:04:43.840
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
and you want to quickly learn the new thing as well.
https://karpathy.ai/lexicap/0011-large.html#00:04:48.080
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So one simple way of doing that is you take the network
https://karpathy.ai/lexicap/0011-large.html#00:04:53.400
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
which already knows 100 types of databases
https://karpathy.ai/lexicap/0011-large.html#00:04:57.720
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
and then you just take the top layer of that
https://karpathy.ai/lexicap/0011-large.html#00:05:02.440
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
and you retrain that using the new label data
https://karpathy.ai/lexicap/0011-large.html#00:05:06.320
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
that you have in the new image database.
https://karpathy.ai/lexicap/0011-large.html#00:05:11.320
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
And then it turns out that it really, really quickly
https://karpathy.ai/lexicap/0011-large.html#00:05:14.760
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
can learn that too, one shot basically
https://karpathy.ai/lexicap/0011-large.html#00:05:17.360
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
because from the first 100 data sets,
https://karpathy.ai/lexicap/0011-large.html#00:05:20.600
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
it already has learned so much about computer vision
https://karpathy.ai/lexicap/0011-large.html#00:05:24.320
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
that it can reuse that and that is then almost good enough
https://karpathy.ai/lexicap/0011-large.html#00:05:27.560
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
to solve the new task except you need a little bit
https://karpathy.ai/lexicap/0011-large.html#00:05:31.880
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
of adjustment on the top.
https://karpathy.ai/lexicap/0011-large.html#00:05:34.240
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So that is transfer learning.
https://karpathy.ai/lexicap/0011-large.html#00:05:38.400
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
And it has been done in principle for many decades.
https://karpathy.ai/lexicap/0011-large.html#00:05:41.280
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
People have done similar things for decades.
https://karpathy.ai/lexicap/0011-large.html#00:05:44.520
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Meta learning too, meta learning is about
https://karpathy.ai/lexicap/0011-large.html#00:05:48.520
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
having the learning algorithm itself
https://karpathy.ai/lexicap/0011-large.html#00:05:51.080
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
open to introspection by the system that is using it
https://karpathy.ai/lexicap/0011-large.html#00:05:55.760
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
and also open to modification such that the learning system
https://karpathy.ai/lexicap/0011-large.html#00:06:01.560
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
has an opportunity to modify
https://karpathy.ai/lexicap/0011-large.html#00:06:06.320
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
any part of the learning algorithm
https://karpathy.ai/lexicap/0011-large.html#00:06:09.680
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
and then evaluate the consequences of that modification
https://karpathy.ai/lexicap/0011-large.html#00:06:12.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
and then learn from that to create
https://karpathy.ai/lexicap/0011-large.html#00:06:16.840
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
a better learning algorithm and so on recursively.
https://karpathy.ai/lexicap/0011-large.html#00:06:21.000
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So that's a very different animal
https://karpathy.ai/lexicap/0011-large.html#00:06:25.680
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
where you are opening the space of possible learning
https://karpathy.ai/lexicap/0011-large.html#00:06:28.480
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
algorithms to the learning system itself.
https://karpathy.ai/lexicap/0011-large.html#00:06:32.440
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Right, so you've, like in the 2004 paper, you described
https://karpathy.ai/lexicap/0011-large.html#00:06:35.480
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
gator machines, programs that rewrite themselves, right?
https://karpathy.ai/lexicap/0011-large.html#00:06:40.160
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Philosophically and even in your paper, mathematically,
https://karpathy.ai/lexicap/0011-large.html#00:06:44.480
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
these are really compelling ideas but practically,
https://karpathy.ai/lexicap/0011-large.html#00:06:47.480
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
do you see these self referential programs
https://karpathy.ai/lexicap/0011-large.html#00:06:52.280
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
being successful in the near term to having an impact
https://karpathy.ai/lexicap/0011-large.html#00:06:55.280
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
where sort of it demonstrates to the world
https://karpathy.ai/lexicap/0011-large.html#00:06:59.360
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
that this direction is a good one to pursue
https://karpathy.ai/lexicap/0011-large.html#00:07:02.960
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
in the near term?
https://karpathy.ai/lexicap/0011-large.html#00:07:07.400
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Yes, we had these two different types
https://karpathy.ai/lexicap/0011-large.html#00:07:08.640
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
of fundamental research,
https://karpathy.ai/lexicap/0011-large.html#00:07:11.320
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
how to build a universal problem solver,
https://karpathy.ai/lexicap/0011-large.html#00:07:13.440
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
one basically exploiting proof search
https://karpathy.ai/lexicap/0011-large.html#00:07:15.800
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
and things like that that you need to come up with
https://karpathy.ai/lexicap/0011-large.html#00:07:22.960
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
asymptotically optimal, theoretically optimal
https://karpathy.ai/lexicap/0011-large.html#00:07:25.520
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
self improvers and problem solvers.
https://karpathy.ai/lexicap/0011-large.html#00:07:30.280
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
However, one has to admit that through this proof search
https://karpathy.ai/lexicap/0011-large.html#00:07:34.160
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
comes in an additive constant, an overhead,
https://karpathy.ai/lexicap/0011-large.html#00:07:40.640
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
an additive overhead that vanishes in comparison
https://karpathy.ai/lexicap/0011-large.html#00:07:44.480
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
to what you have to do to solve large problems.
https://karpathy.ai/lexicap/0011-large.html#00:07:50.760
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
However, for many of the small problems
https://karpathy.ai/lexicap/0011-large.html#00:07:55.160
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
that we want to solve in our everyday life,
https://karpathy.ai/lexicap/0011-large.html#00:07:58.000
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
we cannot ignore this constant overhead
https://karpathy.ai/lexicap/0011-large.html#00:08:00.880
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
and that's why we also have been doing other things,
https://karpathy.ai/lexicap/0011-large.html#00:08:03.280
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
non universal things such as recurrent neural networks
https://karpathy.ai/lexicap/0011-large.html#00:08:08.120
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
which are trained by gradient descent
https://karpathy.ai/lexicap/0011-large.html#00:08:12.160
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
and local search techniques which aren't universal at all,
https://karpathy.ai/lexicap/0011-large.html#00:08:15.400