Episode
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs
1:20:06
Published: Sun Dec 23 2018
Description
Juergen Schmidhuber is the co-creator of long short-term memory networks (LSTMs) which are used in billions of devices today for speech recognition, translation, and much more. Over 30 years, he has proposed a lot of interesting, out-of-the-box ideas in artificial intelligence including a formal theory of creativity. Video version is available on YouTube. If you would like to get more information about this podcast go to https://lexfridman.com/ai or connect with @lexfridman on Twitter, LinkedIn, Facebook, or YouTube where you can watch the video versions of these conversations.
Chapters
The pinnacle of machine learning is recursive self-improvement through meta-learning, where the machine learns to improve itself and the way it learns.
00:00 - 06:51 (06:51)
Summary
The pinnacle of machine learning is recursive self-improvement through meta-learning, where the machine learns to improve itself and the way it learns. This involves modifying the learning algorithm, evaluating the consequences of the modification, and learning from it recursively, with the goal of creating a better learning algorithm.
EpisodeJuergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs
PodcastLex Fridman Podcast
Markus Hutter's universal algorithm can solve the Traveling Salesman Problem in n to the 5 steps, plus O of 1, with the constant overhead for proof search guaranteeing optimality, while the more practical recurrent neural networks and local search techniques are used for solving small problems like speech recognition and machine translation.
06:51 - 19:16 (12:25)
Summary
Markus Hutter's universal algorithm can solve the Traveling Salesman Problem in n to the 5 steps, plus O of 1, with the constant overhead for proof search guaranteeing optimality, while the more practical recurrent neural networks and local search techniques are used for solving small problems like speech recognition and machine translation.
EpisodeJuergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs
PodcastLex Fridman Podcast
Scientists are in search of the shortest program that computes the entire history of the entire universe as it would make the description of the universe much simpler and less ugly.
19:16 - 26:40 (07:24)
Summary
Scientists are in search of the shortest program that computes the entire history of the entire universe as it would make the description of the universe much simpler and less ugly. There is an alternative explanation where everything considered random is pseudo-random, such as the decimal expansion of pi, which looks random but is not.
EpisodeJuergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs
PodcastLex Fridman Podcast
The Power Play approach allows for searching pairs of problems and their solutions or modifications of the problem solver itself.
26:41 - 36:02 (09:20)
Summary
The Power Play approach allows for searching pairs of problems and their solutions or modifications of the problem solver itself. This enables finding new solutions that the old problem solver cannot solve.
EpisodeJuergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs
PodcastLex Fridman Podcast
The idea of creating general problem-solving machines involves developing artificial curiosity systems that go beyond current knowledge to become more general problem solvers and understand more of the world.
36:03 - 49:27 (13:24)
Summary
The idea of creating general problem-solving machines involves developing artificial curiosity systems that go beyond current knowledge to become more general problem solvers and understand more of the world. As part of this process, curiosity is a natural feature of the recurrent network that learns to solve problems in a more efficient manner by creating little subnetworks that stand for the properties of the agent, hand, and other actuators.
EpisodeJuergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs
PodcastLex Fridman Podcast
Controller Model Systems (CMS) provide the opportunity for the learning model of the world to predict potential outcomes using abstract high-level predictions, which are crucial for success.
49:27 - 1:00:07 (10:39)
Summary
Controller Model Systems (CMS) provide the opportunity for the learning model of the world to predict potential outcomes using abstract high-level predictions, which are crucial for success.
EpisodeJuergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs
PodcastLex Fridman Podcast
Despite concerns about machines replacing human jobs, the speaker argues that innovation and human restlessness will always lead to the creation of new jobs, particularly those in the realm of interacting with others through new mediums.
1:00:07 - 1:10:46 (10:39)
Summary
Despite concerns about machines replacing human jobs, the speaker argues that innovation and human restlessness will always lead to the creation of new jobs, particularly those in the realm of interacting with others through new mediums. Even with the introduction of industrial robots, many new jobs were created that required human skills and experience.
EpisodeJuergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs
PodcastLex Fridman Podcast
The possibility of conquering the entire universe through the advancement of artificial intelligence is discussed, as well as the likelihood that artificial general intelligence systems will not desire to have interactions with humans in the future.
1:10:47 - 1:19:52 (09:05)
Summary
The possibility of conquering the entire universe through the advancement of artificial intelligence is discussed, as well as the likelihood that artificial general intelligence systems will not desire to have interactions with humans in the future.