Juergen Schmidhuber on DeepMind, AlphaGo & Progress in AI

In asking AI researcher Juergen Schmidhuber about his thoughts on progress at DeepMind and about the AlphaGo vs Lee Sedol Go tournament – provided some initial comments. I will be updating this post with further interview.

juergen288x466genova1Juergen Schmidhuber: First of all, I am happy about DeepMind’s success, also because the company is heavily influenced by my former students: 2 of DeepMind’s first 4 members and their first PhDs in AI came from my lab, one of them co-founder, one of them first employee. (Other ex-PhD students of mine joined DeepMind later, including a co-author of our first paper on Atari-Go in 2010.)

Go is a board game where the Markov assumption holds: in principle, the current input (the board state) conveys all the information needed to determine an optimal next move (no need to consider the history of previous states). That is, the game can be tackled by traditional reinforcement learning (RL), a bit like 2 decades ago, when Tesauro used RL to learn from scratch a backgammon player on the level of the human world champion (1994). Today, however, we are greatly profiting from the fact that computers are at least 10,000 times faster per dollar.

In the last few years, automatic Go players have greatly improved. To learn a good Go player, DeepMind’s system combines several traditional methods such as supervised learning (from human experts) and RL based on Monte Carlo Tree Search. It will be very interesting to see the system play against the best human Go player Lee Sedol in the near future.

Unfortunately, however, the Markov condition does not hold in realistic real world scenarios. That’s why games such as football are much harder for machines than Go, and why Artificial General Intelligence (AGI) for RL robots living in partially observable environments will need more sophisticated learning algorithms, e.g., RL for recurrent neural networks.

For a comprehensive history of deep RL, see Section 6 of my survey with 888 references:

Also worth seeing Juergen’s AMA here.

Juergen Schmidhuber’s website.

Can Intelligence Explode? – Marcus Hutter at Singularity Summit Australia 2012

Abstract: The technological singularity refers to a hypothetical scenario in which technological advances virtually explode. The most popular scenario is the creation of super-intelligent algorithms that recursively create ever higher intelligences. After a short introduction to this intriguing potential future, I will elaborate on what it could mean for intelligence to explode. In this course, I will (have to) provide a more careful treatment of what intelligence actually is, separate speed from intelligence explosion, compare what super-intelligent participants and classical human observers might experience and do, discuss immediate implications for the diversity and value of life, consider possible bounds on intelligence, and contemplate intelligences right at the singularity.



Slides (pdf):
Slides (PowerPoint):
Paper: M.Hutter, Can Intelligence Explode, Journal of Consciousness Studies, Vol.19, Nr 1-2 (2012) pages 143–166.

Also see:

Marcus Hutter (born 1967) is a German computer scientist and professor at the Australian National University. Hutter was born and educated in Munich, where he studied physics and computer science at the Technical University of Munich. In 2000 he joined Jürgen Schmidhuber’s group at the Swiss Artificial Intelligence lab IDSIA, where he developed the first mathematical theory of optimal Universal Artificial Intelligence, based on Kolmogorov complexity and Ray Solomonoff’s theory of universal inductive inference. In 2006 he also accepted a professorship at the Australian National University in Canberra.

Hutter’s notion of universal AI describes the optimal strategy of an agent that wants to maximize its future expected reward in some unknown dynamic environment, up to some fixed future horizon. This is the general reinforcement learning problem. Solomonoff/Hutter’s only assumption is that the reactions of the environment in response to the agent’s actions follow some unknown but computable probability distribution.


Professor Marcus Hutter

Research interests:

Artificial intelligence, Bayesian statistics, theoretical computer science, machine learning, sequential decision theory, universal forecasting, algorithmic information theory, adaptive control, MDL, image processing, particle physics, philosophy of science.


Marcus Hutter is Professor in the RSCS at the Australian National University in Canberra, Australia. He received his PhD and BSc in physics from the LMU in Munich and a Habilitation, MSc, and BSc in informatics from the TU Munich. Since 2000, his research at IDSIA and now ANU is centered around the information-theoretic foundations of inductive reasoning and reinforcement learning, which has resulted in 100+ publications and several awards. His book “Universal Artificial Intelligence” (Springer, EATCS, 2005) develops the first sound and complete theory of AI. He also runs the Human Knowledge Compression Contest (50’000€ H-prize).

AGI Progress & Impediments – Progress in Artificial Intelligence Panel

Panelists: Ben Goertzel, David Chalmers, Steve Omohundro, James Newton-Thomas – held at the Singularity Summit Australia in 2011

Panelists discuss approaches to AGI, progress and impediments now and in the future.
Ben Goertzel:
Ben Goertzle with backdrop of headsBrain Emulation, Broad level roadmap simulation, bottleneck, lack of imaging technology, we don’t know what level of precision we need to reverse engineer biological intelligence. Ed Boyed – optimal brain imageing.
Not by Brain emulation (engineering/comp sci/cognitive sci), bottleneck is funding. People in the field believe/feel they know how to do it. To prove this, they need to integrate their architectures which looks like a big project. Takes a lot of money, but not as much as something like Microsoft Word.

David Chalmers (time 03:42):
DavidChalmersWe don’t know which of the two approaches. Though what form the singularity will take will likely be dependent on the approach we use to build AGI. We don’t understand the theory yet. Most don’t think we will have a perfect molecular scanner that scans the brain and its chemical constituents. 25 Years ago David Chalmers worked in Douglass Hofstadter’s AI lab, but his expertise in AI is now out of date. To get to Human Level AI by brute force or through cognitive psychology knows that the cog-sci is not in very good shape. Third approach is a hybrid of ruffly brain augmentation (through technology we are already using like ipads and computers etc) and technological extension and uploading. If using brain augmentation through tech and uploading as a first step in a Singularity then it is including Humans in the equation along with humanities values which may help shape a Singularity with those values.

Steve Omohundro (time 08:08):
steve_omohundro_headEarly in history AI, there was a distinction: The Neats and the Scruffies. John McCarthy (Stanford AI Lab) believed in mathematically precise logical representations – this shaped a lot of what Steve thought about how programming should be done. Marvin Minsky (MIT Lab) believed in exploring neural nets and self organising systems and the approach of throwing things together to see how it self-organises into intelligence. Both approaches are needed: the logical, mathematically precise, neat approach – and – the probabilistic, self-organising, fuzzy, learning approach, the scruffy. They have to come together. Theorem proving without any explorative aspect probably wont succeed. Purely Neural net based simulations can’t represent semantics well, need to combine systems with full semantics and systems with the ability to adapt to complex environments.

James Newton-Thomas (time 09:57)
james.newton-thomasJames has been playing with Neural-nets and has been disappointed with them not being thinks that Augmentation is the way forward. The AI problem is going to be easier to solve if we are smarter to solve it. Conferences such as this help infuse us with a collective empowerment of the individuals. There is an impediment – we are already being dehumanised with our Ipad, where the reason why we are having a conversation with others is a fact about our being part of a group and not about the information that can be looked up via an IPad. We need to careful in our approach so that we are able to maintain our humanity whilst gaining the advantages of the augmentation.

General Discussion (time 12:05):
David Chalmers: We are already becoming cyborgs in a sense by interacting with tech in our world. the more literal cyborg approach we are working on now. Though we are not yet at the point where the technology is commercialization to in principle allow a strong literal cyborg approach. Ben Goertzel: Though we could progress with some form of brain vocalization (picking up words directly from the brain), allowing to think a google query and have the results directly added to our mind – thus bypassing our low bandwidth communication and getting at the information directly in our heads. To do all this …
Steve Omohundro: EEG is gaining a lot of interest to help with the Quantified Self – brain interfaces to help measure things about their body (though the hardware is not that good yet).
Ben Goertzel: Use of BCIs for video games – and can detect whether you are aroused and paying attention. Though the resolution is very course – hard to get fine grained brain state information through the skull. Cranial jacks will get more information. Legal systems are an impediment.
James NT: Alan Snyder using time altering magnetic fields in helmets that shut down certain areas of the brain, which effectively makes people smarter in narrower domains of skill. Can provide an idiot savant ability at the cost of the ability to generalize. The brain that becomes to specific at one task is doing so at the cost of others – the process of generalization.

Ben Goertzel, David Chalmers, Steve Omohundro - A Thought Experiment

Ben Goertzel, David Chalmers, Steve Omohundro – A Thought Experiment