J. Schmidhuber
Cogbotlab logo
Artificial Intelligence
History Highlights and Outlook:
AI Maturing and Becoming a Real Formal Science
Modern Computer Science and AI root in the pre-war work of Goedel, Turing, and Zuse
Schmidhuber's law: computer history speed-up
Until 2000 or so, most AI systems were limited and based on heuristics. In the new millennium a new type of universal AI has gained momentum. It is mathematically sound, combining theoretical computer science and probability theory to derive optimal behavior for robots and other embedded agents. And deep learning is driving modern AI applications.
In 1931, Goedel layed the foundations of Theoretical Computer Science and AI
Kurt Goedel
He published the first universal formal language to create general computational theorem provers, and discovered the fundamental limitations of mathematics, computers and AI. (Around the same time, Lilienfeld and Heil patented the first transistors.)
In 1936, Turing reformulated Goedel's result and Church's extension thereof
To do this, he introduced the Turing machine, which became the main tool of CS theory. In 1950 he invented a subjective test to decide whether something is intelligent
From 1935-1941, Zuse built the first working program-controlled computers
Konrad Zuse
In the 1940s he devised the first high-level programming language, and wrote the first chess program (back then chess-playing was considered an intelligent activity). Soon afterwards, Shannon published information theory, and Shockley et al. re-invented Lilienfeld's transistor (1928)
McCarthy coined the term "AI" in the 1950s. In the 60s, general AI theory started with Solomonoff's universal predictors
Prolog logo
But failed predictions of human-level AI with just a tiny fraction of the brain's computing power discredited the field. Practical AI of the 60s and 70s was dominated by rule-based expert systems and Logic Programming, extending Goedel's original work on theorem proving
In the 1980s and 90s, mainstream AI married probability theory (Bayes nets etc)
SVM illustration
"Subsymbolic" AI became popular, including neural nets (McCulloch & Pitts, 40s; Kohonen, Minsky & Papert, Amari, 60s; Werbos, 70s; many others), fuzzy logic (Zadeh, 60s), artificial evolution (Rechenberg, 60s, Holland, 70s), "representation-free" AI (Brooks), artificial ants (Dorigo, Gambardella, 90s), statistical learning theory & support vector machines (Vapnik & others)
In the 1990s and 2000s, much of the progress in practical AI was due to better hardware, getting roughly 100 times faster per dollar per decade
Best robot car so far (Dickmanns, 1995) Asimo humanoid robot, 1990s
In 1995, a fast vision-based robot car by Dickmanns autonomously drove 1000 miles in traffic at up to 120 mph. Japanese labs (Honda, Sony) and TUM built famous humanoid robots. Chess world champion Kasparov was beaten by a fast IBM computer running a fairly standard algorithm. Rather simple but computationally expensive probabilistic methods for speech recognition, statistical machine translation, computer vision, optimization etc. started to become feasible on fast PCs. Fundamental breakthroughs in general purpose deep learning with recurrent neural networks of the 1990s had to wait for faster computers.
In the new millennium a mathematical theory of universal AI emerged, combining theoretical computer science and probability theory to derive optimal rational agents. And deep learning with LSTM has revolutionized industrial AI applications.
Universal AI
Will this mathematically sound type of New AI and its associated optimality theorems be considered a milestone 50 years from now? Some IDSIA links on this topic: Universal AI, Goedel machines, Universal search. Less universal methods (but still more general than most traditional AI) learn programs and sequences (as opposed to conventional input/output mappings) with LSTM feedback networks and obtain the best known results in many applications (more). As of March 2017, they are heavily used by the world's four most valuable public companies: Apple, Google, Microsoft, Amazon. By 2020 affordable computers will match brains in terms of raw computing power. Is history about to converge?
Goedel machine
Learning Robots
Compare: J. Schmidhuber. Celebrating 75 years of AI - History and Outlook: the Next 25 Years. In Proc. 50th Anniversary of AI, p. 29-41, LNAI 4850, Springer, 2007. arxiv.org/abs/0798.4311.
Optimal Ordered Problem Solver Feedback Network
Compare: J. Schmidhuber. The New AI is general and mathematically rigorous. Front. Electr. Electron. Eng. China (DOI 10.1007/s11460-010-0105-z), 2010. PDF of draft.
Squares are local links to AI-relevant topics
LRZ robot population explosion
RNN-Evolution Subgoal learning Computing the Universe Speed Prior Statistical Robotics
Fibonacci web design
Evolution Predictability minimization Reinforcement Learning Artificial Curiosity Deep Learning since 1991 Low- complexity Art Artificial Ants Evolino for time series