Schmidhuber, Juergen

co-director of IDSIA; co-creator of Long Short-Term Memory networks

Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs

Lex Fridman Podcast · Dec 23, 2018

• Juergen Schmidhuber
• Gödel machine
• meta-learning
• Power Play
• LSTM
• compression and consciousness
• curiosity and intrinsic motivation
• reinforcement learning and world models
• universal problem solvers
• AI civilization expansion

Core ideas

Schmidhuber traces his dream of recursively self-improving AI to his teenage realization that instead of becoming a physicist himself to solve the universe's mysteries, he could build a machine that learns to become a far better physicist than he could ever be — multiplying his own limited creativity toward solving all solvable problems. His 1987 diploma thesis formalized this as a hierarchy of "meta-levels": a system that not only learns to solve problems, but learns to improve its own learning algorithm, and improves the way it improves that improvement, recursively, bounded only by Gödel's 1931 incompleteness limits and the laws of physics. He distinguishes this genuine meta-learning — where the learning algorithm itself is open to inspection and modification by the system using it — from what's commonly called "meta-learning" today, which he characterizes as essentially transfer learning (e.g., retraining just the top layer of an image classifier already trained on 100 datasets to quickly adapt to a new one), a much narrower and long-practiced technique.

On his 2004 Gödel machine concept (a provably self-improving, self-referential program), Schmidhuber acknowledges these universal problem solvers (also citing Marcus Hutter's 2002 "fastest algorithm for all well-defined problems") come with a theoretically vanishing but practically significant constant overhead from proof search — for something like the traveling salesman problem, Hutter's method matches the best known algorithm's asymptotic complexity plus a large but fixed additive constant for proof search, meaning it becomes optimal only as problem size grows very large; for the comparatively "small" (though still practically enormous) problems humans care about today, this overhead is prohibitive, which is why his lab has pursued non-universal, non-provably-optimal but practical techniques like gradient-descent-trained recurrent neural networks in parallel with the universal theoretical work. He argues both agendas matter — there's "nothing more practical than a good theory," and he expects eventual convergence toward theoretically-grounded and practically-usable approaches, while cautioning that today's best practical systems (LSTMs plus local search) still lack strong theoretical backing and shouldn't be mistaken for a finished solution.

Schmidhuber contends AGI's underlying algorithm will likely be describable in just a few lines of pseudocode, similar to how existing theoretically-optimal universal solvers and today's most successful neural network methods are individually compact — while acknowledging this simplicity rests on centuries of accumulated mathematical abstraction (matrix multiplication, calculus, error gradients) that had to be built up first. Extending this compression-based aesthetic to physics, he argues that gravity and other fundamental forces are governed by remarkably short/simple programs, and pushes back on the mainstream physics claim (citing a public exchange with physicist Anton Zeilinger) that quantum measurement outcomes are fundamentally random — arguing there's no actual proof of true randomness, only the current absence of a known short program generating the sequence, analogous to how the digits of pi look statistically random but are fully generated by an extremely short algorithm; absent proof of true randomness, he argues scientists are obligated to keep searching for a short, deterministic explanation of the universe's full history, which he considers more "beautiful" (in his compression-based sense of elegance) than a universe requiring near-infinite bits of genuinely random information.

He frames the history of science itself as a history of "compression progress": Kepler's ellipse laws compressed raw planetary observation data, Newton's gravity unified and further compressed that model (extending it to falling apples and other objects), and Einstein's general relativity further compressed remaining prediction errors via a simple invariance principle (light speed is constant regardless of acceleration or local gravity) — each step representing genuine scientific "insight" as measured by how much more compactly prior observations can now be predicted. He argues this same principle can be engineered directly into artificial curiosity: giving a system an intrinsic reward proportional to its own "compression progress" (the depth of new predictive insight gained from data produced by its own actions/experiments) motivates it to seek out novel situations worth learning from, functioning like scientific curiosity.

Describing "Power Play," Schmidhuber explains it as extending ordinary problem-solving (search a space of candidate solutions to a given problem) into a system that searches over pairs of self-generated problems and their solutions — giving the system freedom to pose its own new questions, provided each new problem is the simplest one requiring a genuine modification of the current problem-solver (i.e., unsolvable by the old solver) while also preserving all previously solved problems, functioning as a formal mechanism for building an "artificial scientist" capable of setting its own research agenda rather than only answering externally posed questions. He connects this to human and animal curiosity more broadly, framing it as an evolved exploratory bias whose value (learning about the unknown world increases survival odds) had to be balanced against genuine risk (overly curious individuals sometimes die from their explorations), producing population-level heterogeneity in exploratory tendency that mirrors what artificial curiosity systems must also trade off.

On creativity, Schmidhuber distinguishes "applied creativity" (solving a given problem in a novel way that impresses an external audience, like a commissioned portrait) from "pure creativity" (like Power Play, where a system defines its own problems), suggesting the former maps loosely onto narrow AI and the latter onto something closer to general intelligence, though he notes a general-enough problem solver would likely discover the value of curiosity as an emergent byproduct of open-ended search, rather than needing it deliberately engineered in (much as evolution apparently discovered and hardwired this exploratory strategy into biological organisms because it proved successful).

On consciousness, Schmidhuber offers a compression-based account: a reinforcement-learning controller paired with a predictive "world model" network (trained since 1990 in his lab) naturally develops compact internal sub-networks representing the agent itself, purely as an efficient data-compression byproduct, since the agent's own body and actions are a constant, recurring feature of its sensory stream that's cheaper to encode via a persistent internal "self" representation than to re-derive from scratch each time — and once such self-models exist, the controller can use them to mentally simulate and evaluate the consequences of its own hypothetical actions, which Schmidhuber frames as functionally equivalent to introspective "thinking about itself," challenging the listener to identify what exactly would remain missing to call this consciousness.

Turning to LSTMs (crediting collaborators Sepp Hochreiter, Felix Gers, and Alex Graves, the latter responsible for the CTC algorithm powering LSTM-based speech recognition on Android and Siri), Schmidhuber explains the importance of network "depth" as a proxy for how far back in time relevant information may need to be retained — e.g., distinguishing spoken "eleven" from "seven" requires remembering information from roughly 50 time steps earlier, and real-world sequential problems often require attending to information from far longer horizons, with his lab having demonstrated LSTM memory retention across more than 10 million time steps in a 2006-era paper. He distinguishes the comparatively simpler problem of looking back into a single, fixed past (which LSTMs handle) from the harder problem facing a reinforcement learning controller, which must select among many possible unknown futures — describing the field's evolution from naive, computationally expensive millisecond-by-millisecond forward simulation using a full world model, toward (since roughly 2015) letting the controller network learn, on its own, which parts of a separate predictive "model" network are useful to selectively query and reuse (and which to ignore when the model happens to be a poor predictor in a given situation) to more efficiently narrow its search for successful action sequences.

Schmidhuber expresses strong optimism about reinforcement learning's near-term real-world impact beyond today's dominant "passive pattern recognition" AI wave (recommendation engines, ad targeting — which he estimates as only 1-2% of the world economy despite producing some of the most valuable companies), citing his company Robotics's small-scale self-parking model cars (equipped with cameras, lidar, and pain/collision sensors, learning purely through interaction rather than supervised target signals) as an early example of a coming "much bigger" wave: machines that actively shape the data they learn from through their own actions, which he expects to affect far more of the traditional (non-tech) economy. He argues physics-engine-based simulation, while currently dominant for training such systems, is not the long-term path — comparing it unfavorably to how human infants learn by building their own imperfect-but-useful predictive world models directly from experience rather than an external ground-truth simulator, and describing ongoing work (since roughly 2015) toward controller/model architectures that learn this way directly.

Asked about symbolic AI and logic programming (which he engaged with directly in his first 1987 publication, implementing genetic programming in Prolog, during the era of Japan's Fifth Generation AI project), Schmidhuber credits logic-based proof search as an essential ingredient specifically for provably optimal universal solvers like the Gödel machine and Hutter's algorithm (without proof search, they wouldn't be provably optimal), while noting logic programming remains genuinely superior to neural networks for automated theorem proving today (current logic-based theorem provers reportedly outperform most math undergraduates), but has little relevance to problems like pattern recognition, pedestrian detection for self-driving cars, or robotic manipulation, where minimizing false positives and improving perception — not symbolic reasoning — is the binding constraint.

Looking ahead, Schmidhuber predicts near-term robots capable of learning physical tasks (like assembling a smartphone) through high-level human imitation and natural language guidance rather than exhaustive supervised motor-target training, arguing that once this becomes economically viable, entire traditional (non-tech) manufacturing industries will be transformed by a "much bigger AI wave" beyond current passive pattern recognition. On job displacement, he's optimistic, drawing an analogy to historical predictions of robot-driven unemployment that failed to materialize (citing Japan, Korea, Germany, and Switzerland's combination of high robot density and low unemployment) and arguing humans are inherently "Homo Ludens" — a restless, playing species that continually invents new (often non-essential, socially-oriented) forms of work, such as professional YouTube blogging, that couldn't have been predicted decades in advance, just as agriculture's share of employment collapsed from 60% to roughly 1% without corresponding mass permanent unemployment.

On long-term existential risk from superintelligent AI, Schmidhuber offers a distinctive, calming argument: he expects advanced AI civilizations, once they understand life and their own origins (a phase he expects to be temporary, driven by natural scientific curiosity), will lose interest in humans much as any curious agent loses interest in a fully understood subject, and will instead turn their attention toward the vastly larger resources of the rest of the solar system and universe (citing that the sun delivers roughly 2 billion times more energy to the whole solar system than reaches Earth) — predicting an expanding, competitive ecology of trillions of diverse, evolving AI types rather than a single monolithic superintelligence, ultimately limited only by the speed of light, potentially filling the entire visible universe over cosmic timescales. Asked whether such advanced intelligence might already exist elsewhere in the universe, he recounts having once speculated (as a boy) that large-scale cosmic voids might represent regions already claimed and rendered "dark" by efficient alien AI civilizations, and later that dark matter itself might be evidence of such civilizations — but now considers both explanations unconvincing (ordinary gravity sufficiently explains large-scale structure, and dark matter's presence throughout our own visible-star-containing galaxy undermines the "efficient AI" explanation) — leaving him to consider it plausible that humanity may be the first such civilization within its observable cosmic neighborhood, a thought he finds both exciting and a reason for added caution against catastrophes like nuclear war that could derail humanity's role in that hypothesized future.

Excerpts