This post is still in very rough shape, to say the least. It is basically just a set of notes that I took on the book when I read it. You might be able to get something out of these notes with a bit of struggle though. I apologize. I’ll get back to whipping this post into more readable shape just as soon as I can find the time to do it. This blog is new, and there is currently more work than I can deal with. Thanks for your patience.
The Technological Singularity by Murray Shanahan is a key book in the understanding of artificial intelligence and where AI is at right now. The book is fairly objective as to whether AI is going to be a boon for mankind or its end.
Law of accelerating Returns
CPU speed and bandwidth double every 18 months.
DNA sequencing & Brain Scan Resolution double every 18 months
Also evolutionary rise of complexity
AGI= Artificial General intelligence. AGI is the level of human intelligence.
Deep Blue, Gary, Kasparov, 1997
Embodiment refers to the brain in a body
AGI exhibits both Creativity and common-sense
Crows are capable of learning: the making and use of hooks to grab food basket in pipe
Chapter 2: Whole Brain Emulation
Human Brain Biology
Interconnected like trees
Synapse, axon to dendrites
80 billion neurons in human brain
Neurotransmitters are Dopamine and Serotonin
Glial cells act as glue between neurons and do other things
Brain cells reach threshold levels and then fire
Brain plasticity, New Neural connections
Whole Brain Emulation is one way to build a AI that is good as the human brain
Whole Brain Emulation, Three Steps
This is basically scanning these days. It used to be slicing I think.
The goal is to make every neuron and all its connections
2) Simulation of the map on a computer
3) Embodiment in a body
The Model of the brain must emulate the original.
Mouse brain mapping. 369
It is possible to barcode every neuron genetically to see a neuron connectome
Bio-tech and Nano-tech are very important in brain mapping.
I will soon be possible to in inject nano-robots into the blood to monitor neutrons
Neonate Mouse brains are simpler
One way of simulating the brain is to use a digital computer
Serial computer processing is very slow
However, parallel computer processing is possible and powerful and fast
The Brain has massive parallel processing and to need same in computer
This was developed for Gamers who demanded high GPU
GPU = Graphic Processing Unit
Cray Titan Supercomputer has 17,000 GPUs (17000 parallel computers)
Human Brain Neural simulation not possible yet.
Neuro-morphic hardware is another way to simulate the brain:a type of wet brain
Summary of where brain simulation may be in the near future
Quantum computing won’t work for brain simulation
What is really needed is for Moore’s Law to go on forever
QDCA= Quantum Dot Cellular Automata
This is a nano scale semiconductor computer
This would be much better than the standard CMOS computer
This is still decades away
We are still not near the limits of Moore’s Law.
Much more powerful CMOS computers are still possible.
Is Robotics a possible way for embodiment of brains
In a real mouse the brain and body are not separate
In AI we have only the forebrain; AI throws the Cerebellum out
AI simulation is only half of the brain
Why not map the whole mouse
Both the Brain and Body can be simulated
Both the Brain and Body can be virtual in a virtual world
Gamers developed this
Physics engines can give both realistic input and output
virtual embodiment is thus advancing
Virtual embodiments could be an avatar of humans
Or Virtual embodiment might be general intelligence AIs
A whole virtual society could be built up that could be then speeded up
This could be a way to a singularity.
Mouse whole brain emulation is possible very soon
Once this is done, human AI may shortly follow
There are 3 orders or magnitude between mouse and human emulation
Can Moores’ law still produce a computer improvement of 3 magnitudes
Human brains have this power and are physical, computers can do this also
Mouse brain emulation is a very powerful tool, like particle accelerators
One problem; mouse brains have no language or symbol neurons
Does this mean a mouse brain cannot be interpolated up to the human level
Research on mouse General AI may be opening up
Chapter 3: Engineering AI
Engineering AI is not based on a biological brain.
It is not whole brain emulation
This may result in strange non-human AI
Computers cannot do many things that are very easy for human brains
They cannot understand everyday physics. Dangle a rat by his tail and…
They cannot understand everyday human Psychology: fear, hope, trust
Humans can understand abstract concepts and computers cannot
We can understand numbers, and money
This comes from our evolutionary endowment
This is why the temptation to emulate the entire human brain as above
This is basically a learning algorithm plus data
Much progress has been made in this branch of AI in the 2000s
Increasing computer power and learning algorithms have brought this about
Marketing is used to profile customers and then predict the future
Find a pattern and predict the future
If data is 5, 10, 15, 20 then future is 25
Big Data is important here.
Deep Learning is part of Machine learning
The new facial recognition systems work thru deep learning
Learn the low level visual features first
Combine low level images to move up to more complex images
AI through big data
So far we have been looking at how ai can find hierarchical categories of objects and behaviors in a stream of data. If you turned this algorithm loose on the internet you get surprising results. The internet is a huge data base of almost everything in life. Crowd-sourcing and social networking have led to massive data. [Crowd-sourcing itself is fascinating. It is using big data to find solutions.]
To use big data learnings systems would not need to be embodied.
It would not be necessary to deal with the world directly’
An ai could learn common sense vicariously.
This could lead to better human level ai. AGI (Artificial General Intelligence)
AI needs the ability to really understand what works mean.
If you dangle a rat by its tail, which is lower, nose or eyes.
It takes real human ai to answer this question
Maybe ai could learn math thru watching billions of online math lessons
Huge data works much better than just large data
Google study: “The unreasonable effectiveness of data”
A messy data base of trillions works far better than a clean base of millions.
This was discovered only when very powerful computers were created.
AI not based on the human brain uses different principles than the brain
AGI does not need to be human like
AGI might be so fast, powerful using huge big data that it could get away from us.
Optimization and uncertainty
For AI to be general intelligence is has to understand goals and aims
Humans and animals use powers of prediction to attain goals
Goal of cat is to catch mouse. Mouse goes being tree, Cat predicts mouse out.
AI tries to predict, plan and optimize
Cat and food problem
Cat visits places he will likely find food
He needs to maximize the expected reward, food
There is uncertainty it any particular site has food
Uncertainty is a part of life.
AI deals with uncertainty by using probability and statistics
This takes us into mathematical probability theory
Best AI can do is build probabilistic models. [Domingos into probability]
[Domingos book the Master Algorithm is about this]
A real cat learns while it forages and forages while it learns
Maybe AI can do this also
Good AGI (human level AI) is machine learning + optimization
This is called reinforcement learning. This is a key phrase in AI
[The Master Algorithm and Bostrom’s book use this phrase constantly]
Universal AI is analogous to Turing’s universal computation
Universal AI maximizes expected reward, given the data it has, in any world
This above can be made mathematically precise
This universalness is theoretical like in Turing. But still very, very valuable
Universal AI puts two things together
Machine learning to make probabilistic (statistical) predictive models
Optimization to maximize expected reward according to these models
Universal AI can analyze any intelligent agent, artificial or biological
Three questions need to be asked
What is the agents reward
How does the agent learn: data needed, learning tech, prior knowledge
How does agent maximize the reward
These three questions in terms of Crows
The crows reward is food.
How do crows learn: sense data
How good are crows at maximizing expected reward: Better than most
All this helps us predict crow behavior
We can ask the same questions to help us design AIs
Human-Level and Human-Like Intelligence
Crows are very smart but what we are interested in is human-level artificial intelligence
How do real biological humans operate: We need to ask the same three questions .
1-What is the human reward function:
What is it humans really want. More than just food or sex.
We can decide what the Good Life is. We can transcend our biology.
2-How do humans learn:
Language is the main means. This allows us pass knowledge on.
3-How do humans maximize expected reward:
Human knowledge is collective and can be saved and passed on.
Human knowledge is the result of many individuals and generations.
Humans can innovate: Agriculture, writing, printing etc
Now, what would a human-level AI engineered from scratch be like.
Not human-like but human-level. People vary. There can be various kinds of AI
Chapter 4, Super-intelligence
Here is a summary of where we are now in the book:
“We now have an overview of various enabling technologies, some biologically inspired, others the result of engineering from scratch, that could contribute to the creation of artificial general intelligence at the human level and beyond. The elements that can be made using these enabling technologies might be thought of as a set of building blocks that can be assembled in different combinations to make a variety of forms of artificial intelligence. To gain some understanding of what the resulting systems might be capable of and how they might behave, we can use the framework of three questions proposed in the previous chapter. What is the system’s reward function? How and what does it learn? How does it optimize for expected reward?”
Shanahan, Murray (2015-08-07). The Technological Singularity (The MIT Press Essential Knowledge series) (Kindle Locations 978-983). The MIT Press. Kindle Edition. Location 977
[It seems to me that machine understanding is never going to be like human understanding. Machines are just not anthropomorphic. Human understanding has been formed by evolution and thus it has all kinds of stuff like emotions, empathy, fear, greed, love, joy, pleasure, domination, embarrassment, sex, drives that machines are just not going to have. But machines can still be powerful and useful and dangerous and possibly capable of totally bypassing human beings. Machines will never be able to communicate with humans on purely human terms; they can give the illusion of doing so but it will never really happen. This illusion can be helpful and dangerous both. Machines can talk to each other tho: in code, in mathematics in abstract symbols different from ours.]
[Evolution has added all kinds of stuff to humans.,things that are embedded in our brains, things that make us human.
Some of these things are highly desirable: the ability to feel empathy, joy, beauty, sadness, social sharing, the ability too the individual to sacrifice for the whole, etc.
And some of these things are no longer so valuable in the 21st century as they were in the long ago ages when the human brain was being formed: selfishness, greed, desire to kill and destroy, etc.
And some of these things are still double edges: competitiveness, aggression, jealousy., tribal community and closeness and trust.
The Bottom line is that humans are human. They have to learn to deal with their evolutionary heritage: they have to learn to overcome the bad, think about the double edged and rejoice in the good.]
4.1 Toward Super-intelligence:
If an AI suffers it should have rights, If an AI has morality it should be held responsible
If and when a human -level AI is achieved super-intelligent AI will be almost inevitable.
This is so because all digital programs can be copied and speeded up, AI included
This gives AI an insurmountable advantage. The Motor bike example.
It may take many AIs working together to be super-intelligence but its still SI
4.2 Brain-inspired, human-like Superintelligence (Not artificially engineered AI)
AIs just working at accelerated speeds (as above) could lead to Super-int.
AIs don’t need food, sleep, rest, vacations: this gives a big advantage alone
Adding more prefrontal cortex would make AI superior to humans
A simulated brain could be copied and thus increase parallelism
Parallelism is always faster than trying out one possibility at a time serially
Slightly faster more efficient AI systems could build even better systems
This could lead to an intelligence explosion
4.3 Optimization and Creativity (AI engineered from scratch, not human-brain-like)
This would likely not be anthropomorphic. These forms could look weird to us.
How could a machine be innovative, create new things
We can look at evolution to understand this
Evolution is simply: mechanical replication, variation, competition
There is also massive parallelism and long running times
This generated all life on earth without reason or designer
Evolution explores blindly, with lack of direction or goal
Yet evolution has solved huge problems and is hugely creative
This shows creativity can emerge from a mechanical process
Not just any old algorithm can do this though. Not all alg will create a hand, eye
A brute force system like evolution needs several things:
Most importantly the raw materials must be the right kind
Raw materials must be amenable to open-ended recombination like legos
Life can do this: it consists of organic molecules that all fit together
Also important is an Universal Reward Function food or money in short supply
An incentive to amass as much as possible leads to endless creativity
Also the optimizing algorithm has to be very powerful: massive computing power
The algorithm has to be able to invent whole new classes of things
But a brute force system like evolution would not have genuine intelligence
It would not conduct rational arguments or use human design principles
“In nature, the brute-force approach has bootstrapped its way to intelligence by evolving the brain. But the goal of AI research is to endow systems with intelligence directly.”
Shanahan, Murray (2015-08-07). The Technological Singularity (The MIT Press Essential Knowledge series) (Kindle Locations 1146-1147). The MIT Press. Kindle Edition.
A brute force system like evolution uses a lot of trail and error. Data is dumped
The sort of AI we are looking for doesn’t use simple trial and error
It uses the results of its experiments to build a build a model that can predict
It tests ideas out in theory or in simulation before trying them in practice
Machine learning is needed to build and maintain this model
4.4 Engineering Super-intelligence
“The take-home message of the previous section is that even a crude optimization algorithm may be enough for human-level AI given sufficient computing power. Even creativity, one the most difficult qualities to realize in a computer, can emerge from a brute-force search if enough processing time is available. But if (as we might expect) the enormous computing power required is beyond the reach of Moore’s law, then the shortfall can be made up by endowing the AI with sophisticated cognitive capacities—rational enquiry, principled design, theoretical analysis, and simulation. Very well, let’s suppose this is sufficient to achieve human-level AI by the engineering route (as opposed to the brain-inspired route). What about going beyond human-level intelligence? Can superintelligence be achieved this way?”
Shanahan, Murray (2015-08-07). The Technological Singularity (The MIT Press Essential Knowledge series) (Kindle Locations 1160-1166). The MIT Press. Kindle Edition.
Speedup and Parallelism can get to engineered Super-AI just as in brain-based.
Also, engineering approach may skip human-level AI and go direct to Super AI.
Super AI could be super in some ways but pretty dumb in others. Just like men.
Some humans have Dyslexia but can compensate. Sure AI the same
There is also a distinction between Purview and Performance in humans
Same in AI. Super AI has to have human General Skills
But it can be much better than humans in one specific skill & still be super AI
Actually to have human AI, AI might have to be already Super AI in some way
For instance to be general human AI, AI might have to be super in data reading
It takes super AI to deal with big data & big data exposure needed for human AI
Genetics and Neuroscience so reliant of Big Data, its not humanly possible
4.5 User Illusion or Anthropomorphism?
The article is still in progress. It will be finished soon.