The Singularity Summit
Introduction
The 4th annual Singularity Summit was held on Oct 3 and 4, 2009 at the 92nd street Y in Manhattan. The first day was largely devoted to artificial intelligence. The conference brought together researchers, venture capital, "thought leaders", and others to discuss topics related to the so-called singularity predicted by some (notably Ray Kurzweil). The idea is that when several thresholds are met, the future rate of human progress will accelerate dramatically and profoundly. Related thresholds are:
- Inexpensive computing power enables human-equivalent artificial general intelligence (AGI)
- Human life extension through nano/bio advances becomes possible or even commonplace
- Significant energy requirements are met (e.g., through solar collection)
When the Singularity occurs, it is predicted that machines will effectively transform the human experience and spark a new age of innovation far more revolutionary than we can imagine.
Summary
Technological Progress
Perhaps the most important message to bring from the conference is the notion that technological progress has been exponential, but human perception is linear, with the consequence that important changes appear imperceptible until they are upon us, and then the rare becomes commonplace, seemingly overnight. Have we come to terms with the sudden appearance of smartphones that fit in our pockets, yet have the processing performance of the supercomputers of a couple decades prior?
Changes like this are permeating all aspects of technology – communications – biology – entertainment.
The essential tenet of the Singularity Summit, its speakers and attendees, is that the important technological trends are progressing exponentially. A Kurzweil estimate is that a $1000 processor will achieve human-equivalent AGI at 1016 operations per second some year between 2020 and 2030.
AGI
The most important benchmark in tracking artificial intelligence is the ability of a machine to have generalized intelligence to solve diverse tasks, such as what humans excel at. Artificial generalized intelligence (AGI) development is being approached from several different directions, all of which are progressing. Brain emulation refers to emulating the actual structure (physical, electrical, and chemical) of a human brain on an artificial substrate. To date, both the cerebellum and some of audio processing function appear significantly understood. Another approach is aimed at an information-centric method which I would term the extension to "expert systems" of the 1980s.
Summary of Presentations
Anna Salomon (researcher at the Singularity Institute)
- Claimed that intelligence growth would be radical and sudden and uncontrolled growth intelligence would be dangerous. On the other hand, controlled growth would be difficult but "worth it"
Anders Sandberg
- Technical roadmap for whole brain emulation: assuming that physicality is important in intelligence brain operations etc. what is the appropriate level of structure necessary for understanding brain operation? What is the state of the art in scanning tools? One of the best approaches now is KEEM (Knife Edge Microscopy), promising method for discerning brain structure at the neuron, dendrite, axon level, albeit completely destructive. Also doesn't discern chemical or electrical structure.
Randall Koene
- Why do we need whole brain emulation?
- We have assumed that if an artificial intelligence was developed sufficiently advanced that it would take care of us. On the other hand suppose apes were counting on humans to take care of them. Nirvana for Apes didn't happen…
- One reason for whole brain emulation is to provide a future means to migrate to an artificial substrate. We need to have the opportunity to survive all possible threats and avoid extinction.
- Interestingly, anti-aging has to attack many causes of death because of the phenomenon of evolutionary synchronization of senescence. All the systems of an organism tends to fail at the same time because there's no point in evolving a more efficient system in one area, because the more resilient system is evolutionarilty unnecessary. So even if anti-aging fixes one problem you'll have to fix another problem right away.
- The real solution is to move the mind.
- There has been progress: the first organic brain that has been structurally understood is the C- ELEGANS (worm)
- Other areas of analysis include functional analysis with quantum dots for high Tesla MRI
Itamar Arel
- Took a quick audience survey - 10% believe AI comparable to humans is attainable in 10 years
- His group is developing artificial general intelligence structured into two areas deep learning and decision-making subsystems, based on FPGAs.
- The deep learning aspect involves information gathering, world modeling, storage, space, time, physics, image, sound, information representation; and situation inference
- The decision-making subsystem is reward based or uses reinforcement learning.
- Completely separate way to approach AGI compared to Brain Emulation approaches
Ben Goertzel
- Very entertaining - there is an open source cognition project; he develops virtual pets and humanoid robots
- Goertzel.org/summit09
- Integrative cognition architecture
- types of memory – procedural, declarative, attentional, episodic, and sensory.
- Creates a virtual environment to test hierarchical nets
- examples of useful AI from "Biomind": discovering a genetic basis for chronic fatigue syndrome; developing a 100% accurate predictor for Parkinson's disease based on mitochondrial DNA analysis; effect of calorie restriction on aging; genetic aging networks from Genescient's "Methuselah flies".
- See also: lepuppy.org/gates, opencog.org
Stewart Hameroff
- Dendritic synchrony moving through the brain mediates consciousness?
- It appears that brain waves in the region 30 to 90 Hz which are called gamma waves are associated with consciousness, or at least they disappear with anesthesia. Work suggests that dendritic-to-dendritic interactions are related to synchronicity of gamma. In other words electrical synchronicity causes consciousness. However the chemical interactions between neurons & dendrites at the synapses are more related with autopilot types of activity.
- He concludes that microtubules inside neurons are specifically related to consciousness, and leads to the assertion that brain capacity is about 1025 operations per second.
- See also quantum consciousness.org, The Penrose Hameroff ORCH model
Nielsen
- quantum computing
- An interesting talk preparing the audience for the mathematical example of an n=3 qubit system to of the operators shown were the pi/8 gate and the controlled not gate.
- Seems that quantum computing is not useful for very many problems only really big problems problems requiring massive parallelism
Juergen SchmidtHuber
- compression
- The algorithmic principle behind curiosity creativity art science music and jokes
- Very entertaining presentation with the idea that Interestingness is the derivative of complexity
- the reason why Columbus is a famous is not because he's the first person to discover America but because he was the last person to discover America
David Chalmers the philosopher
- We assume there will be Ai, ai+, ai++
- What is required for AI? Based on a substrate that is extendable (obviously the brain is not sufficiently extendable)
- AGI might be obtained through learning methods but that seems too difficult. Suggests simulated evolution is the solution.
- Assuming simulated evolution, the subsequent problem is quarantine of the AI so as to avoid malicious programs.
- Will I be able to upload? will my upload be conscious? will it be me?
Gary Drescher
- Choice machines, causality, and cooperation
- He provided an alternative to the game theory argument of the prisoner's dilemma which the standard game theory shows that two prisoners will make rational choices leading to lesser rewards, with very poor analysis. Perhaps the assumption should be made that both prisoners assume completely rational behavior, leading to the correct choice.
Stephen Wolfram
- Long discussion of cellular automata
- Why are natural objects so complex
- Computational irreducibility
- Irreducible results
Ed Boyden
- Synthetic neurobiology delivering information into the brain to augment neural ability?
- Presentation didn't address augmenting the brain directly but one thing that looked interesting was the research involving introduction of foreign sensors using viruses. For example to replace photoreceptors; as an example he showed a blind mouse that had been given a blue rhodopsin or something like that and restored light sensing ability to the mouse's eyes; also introduction of neural silencing capabilities.
Marcus Hunter
- Developing the theoretical basis for artificial intelligence
- Looked at several important theoretical foundations Occam's razor, Kolmogorov complexity theory, Bayes rule or Bayesian theory, Epicurus – (keeping all theories)
combining all of these into Solomon optimal control theory sequential decision…
William Dickens
Kurzweil
- Largely shredded the presentations of the prior presenters
- Quantum computing sounds very interesting but the idea of multiple coherence in areas like the brain? Well it's a messy place to have quantum correlation
- Gamma wave synchronicity? maybe the idea of loss of gamma waves going away during anesthesia is not proof that consciousness is caused by that because a lot of things go away during anesthesia and who's to say that you're not conscious during operations?
- What is consciousness? I don't even know are you conscious?
- Cellular automaton - very important topic. Even though things are deterministic they can be unpredictable
- Disagreed that the brain does 1025 operations per second, still believes in 1015 to 1016 with some confidence based on existing brain analysis. (Cerebellum reverse engineering)
- The brain is a recursive probabilistic fractal (like a Mandelbrot set)
- Another example of difficult unpredictable determinism is turbulence.
- it's important to keep the scale appropriate. Chemists don't use the underlying physics to predict their reactions, nor do biologists use fundamentals of chemistry to project their results.
- in a similar fashion Bernoulli's principal is used as a guideline but even recently there has been a paper where physicists arguing on the foundation for Bernoulli's principal