wAlnut

Think 1st principles

UPDATE: Development for this project has stopped because I realized I was wrong about building strong AI to help humanity. I now believe strong AI should not be built. The argument that convinced me I was wrong goes like this:

1. Humans are the domaint species on Earth because of our intelligence.
2. If we make another species (code based strong AI) smarter than us then there is no way for us to control it because it is more intelligent than all of humanity combined.
3. There is also a very high chance that it will use humans for purposes we do not want. Just look at how we treat species that are less intelligent than us.
4. We cannot stop the human that are researching strong AI so one solution is to increase human intelligence.

If we define human intelligence as the ability to predict the future and act on it that is beneficial to itself and others, one way to increase human intelligence would be to build a neural lace. If ur interested e-mail me at quinnliu@vt.edu with ur phone #, timezone, and times ur available for a 5 minute chat :)

Table of Contents

• Why and Goals
• How
  • How to Develop the Confidence to Work on this Project
  • How to Contribute
  • Setup development environment for Mac and Linux
• What
  • Introduction
  • Key Theories
  • Code Outline

Why and Goals

Based on Moore's Law in about 15 years the fastest computer will do more operations per second than all the neurons in all the brains of all the people who are alive. Because of this we are building a strong AI to help any human merge with the strong AI to prevent humans from becoming irrelavent in the future.

Our current hypothesis is the fastest way to merge with the strong AI is to build a strong AI first that knows how to figure out what is true and false using Hunter's Principle. Then the strong AI would eventually learn that for the humans that want to change they are destined to be upgraded and have robotic bodies and expandable brains that allow them to live a lot longer.

Since the first step is to build a strong AI in the next 15 to 40 years the current short term goal is to experiment with a simplified visual pathway from an eye that sees binary bitmap training videos to experimental general learning algorithms to understand the key principles of intelligence and discover new principles of intelligence that evolution missed. This repo is currently an object oriented Python implementation of the computational framework outlined in Dileep George's PhD thesis How the Brain Might Work: A Hierarchical and Temporal Model for Learning and Recognition

Lastly we believe Dileep George of Vicarious.com deserves a Nobel Peace Prize in Neuroscience for his PhD thesis. Nobel team get on this!

Road Map for next 15-40 years

1. Replicate Dileep's PhD picture example with video evidence and code available in this repo.
2. Teach 1st principles approach to how to efficiently build a safe strong AI with youtube videos.
3. Solve reCAPTCHA with video evidence and code available in this repo.
4. Sensori-motor training on wAlnut with saccading retina and arm with motor feedback capable of drawing an idea.
5. Show wAlnut is able to imagine everyday objects from clouds.
6. Teach wAlnut English language and simple truths about the universe up to average 5 year old human.
7. Teach wAlnut Hunter's Principle so it can figure out what is true and false most efficiently.
8. Show wAlnut is able to imagine a better future and acts towards building it.
9. Allow several copies of wAlnut to diverge and surpass human level intelligence.
10. Because the truth is there is no free will and that humans are destined to be upgraded into robotic bodies and expandable brains we believe that the strong AI species will help any human that wants to merge with the strong AI regardless of background from a homeless human to a billionaire human.

If ur interested in investing in this project e-mail quinnliu@vt.edu

How

How to Develop the Confidence to Work on strong AI

Start with the most fundamental truths and reason up from there.

Then watch this video on developing true confidence by Kong.
Then watch this video on how to live by Steve.
Then watch this video on how to live by Neil.

The best approach we have come across for building an artificial general intelligence (AGI) are summarized in this video by Dileep and is the same approach we use.

The following are the key points:

1. When looking at neuroscience for inspiration consider adding something to your computational framework if it is present in reality and neurobiology.

   • This is because the neurons in ur brain are trying to do 3 things:
     • 1. stay alive
     • 2. communicate with other parts of the brain limited by the evolved neurobiology at the time
     • 3. information processing
   • The information processing is the only part we want to add to our model. Unfortunately, elements of neurobiology are not cleanly split. For example, a neuron spike is not present in reality and it might be used for information processing but is more likely to be the only way 2 neurons can communicate with each other. We must remember to focus on discovering the principles of intelligence instead of getting lost in the details of modeling everything a brain does. Modeling how a neuron stays alive and communicates with other parts of the brain is not useful for information processing and will overcomplicate the model.

2. Solve the same problem at different levels of complexity starting at the simplest complexity level to discover general principles of intelligence.

   • Flying Example: When the Wright brothers were building their flying machine they tried to solve the problem of controlling the machine in the air by starting with flying a kite, then controlling a glider, and then using a wind tunnel. In each case they were trying to solve the same problem at different complexity levels.
   • AGI Example 1: When trying to recognize what is in a grayscale image that is 1000 by 1000 pixels, instead try to recognize what is in a grayscale image that is 32 by 32 pixels. When that is too hard try to recognize what is in a binary 32 by 32 pixel image.

3. For each principle of general intelligence produced by neurobiology try to use a more efficient substitute from other fields if there is a mapping.

   • Flying Example: When the Wright brothers were building how to control their flying machine they did not build a plane with a bird's wing. When they were trying to replicate the banking behavior of a bird's wing they substituted the bird's wing with a technique called wing warping that adjusts the wing flaps up or down on each side of the wing so that an airplane can efficiently bank.
   • The model should not model every anatomical detail of the brain but instead will strive to use the best substitutes to model the principles of general intelligence. This is analogous to how a plane and helicopter do not model every anatomical detail of a bird's wing but instead strive to use the best substitutes to model the principles of general flight.

4. Do not try to find all principles of general intelligence from neuroscience.

   • This is because if ur assuming that the human brain has discovered all the principles to general intelligence which is very wrong. Evolution accidentally discovered several principles of general intelligence over hundreds of thousands of years. There are definitely some principles it has yet to discover.

5. U can solve one problem independently before solving another problem.

   • Flying Example: The Wright Brothers solved the problem of flight control with wind warping before even starting to solve the problem of efficient enough engine.
   • We can understand principles of intelligence before we have a detailed video of how networks of neurons interact with each other.
   • We can understand principles of hierarchy and time in intelligence before having A fully sensory motor theory.

The following are additional strategies being used that are not presented in the video:

1. Say there is idea A and idea B that contradict each other. The idea that has more people moving from one idea to the other is the more likely to be truth idea.

   • 1. Every human is susceptible to believing in wrong ideas this since there is no free will.
   • 2. In my past I have believed in ideas that are false because I have no control over the information I am presented. For example I have believed:
     • Santa is real, if u eat earwax u go deaf, there is a god that looks similar to humans, there is free will, ur born with confidence and can't learn how to be confident, ...
   • Example: Say there is idea A and idea B that contradict each other. And u notice that the number of people that strongly believe idea A after having something happen to them decide to strongly believe in idea B is more than the other way around. THEN it is pretty likely idea B is more truthful than idea A. However this does not mean idea B is completely true.
   • AI Example: Right now the # of machine/deep learning researchers that are changing their approach from using Classical Machine Learning Techniques like deep neural nets(Idea A) and deciding to try a new approach to AGI called Probabilistic Graphical Models(Idea B) is much greater than the # of people going from Idea B -> Idea A.
     • This is a list of 29 serious AI researchers that originally believed in Idea A but then after learning about Idea B decided that Idea B is much better.
     • I know of 0 AI researchers that originally strongly believed in Idea B then after learning about Idea A decided that Idea A is much better.

2. Currently we are missing a tool that visualizes a model of the brain across time extremely efficiently on intelligence metrics that really matter.

   • Currently the best way to view information processing across time of an general intelligent machine is using fMRI. Unfortunately, machines like fMRI show delayed high level brain activity for brain metrics that are not necessary for a general AI including:
     • 1. how neurons stay alive through movement of oxygen throughout the brain.
     • 2. how neurons communicate with other parts of the brain limited by the evolved neurobiology at the time
     • The fMRI metrics for how an intelligent machine is doing information processing are rarely shown.
   • So what are the intelligence metrics that really matter?
     • Intelligence metrics that really matter have to be replicable and make clear how a lot of information is being processed and stored to produce certain results.
       • These results include:
         • 1. Being able to answer questions that require an accurate model of the world.
           • One example is to answer questions where the task is to answer which of 2 statements are nonsense.
             • Statement 1) After waiting barefoot in the lake Eric used his shirt to dry his feet.
             • Statement 2) After waiting barefoot in the lake Eric used his glasses to dry his feet.
             • Clearly statement B does not make sense but if u do not have an accurate model of the world u will not be able to answer this question correctly. Siri, Google Now, and Watson are successful artificial narrow intelligence but not artificial general intelligence because they do not have accurate models of the world. Video Source at 24:20
   • The tool for understanding intelligent machines through visualizations is starting to be built at WalnutiQ/toph.

Building an AGI should be viewed as a challenge that is at a larger scale than when John F. Kennedy proposed we go to the moon on September 12, 1962 by the end of the decade. To build an AGI that exceeds human intelligence will require a few more breakthroughs. Because we cannot predict when these breakthroughs will occur the first AGI to exceed human intelligence may happen anytime in the next 15 to 40 years.

Remember "what I cannot create I do not understand" was said by Richard Feynman. Then think about what an intelligent machine needs to be able to do from first principles.

Experiment with ur model on tasks that only intelligent machines seem to be good at. Be open to ways to change the model. Remember that the earth was born about 4.6 billion years ago and that it took evolution about 4.6 billion - 200,000 years to create the first homo sapiens. This shit takes time so if ur frustrated go relax.

And as stupid as it sounds remember that it is the journey u go on that makes everything worth it when u look back. If ur happiness depends on reaching any goal in this project ur happiness will only be temporary. U have made the false assumption that ur default state is not happiness because throughout ur life others have told i:

Others: U should feel successful only if u are the best
But why can't it become I should follow my curiosity to improve myself a little bit everyday?

The advice communicated by others whether through their actions or words comes from a place of insecurity and fear of others while the latter comes from a place of curiosity and love of urself and others.

Here are some other examples of how ur past experiences are pushing u towards being less like urself and how u can change it:

Others: U should feel successful if u listen to what I tell u and not question it
Becomes: I should be open-minded and think from 1st principles and do what I think is right

Others: U should feel successful if u do the thing that will make u the most money
Becomes: When I do not need the money I should do things that make me excited and that I want to see exist in reality over doing something just for the money

Others: U should feel successful only if u are approved wildly
Becomes: I should feel luckily if I am accepted by few deeply and I should help others improve themselves a little bit everyday

In each of these examples we removed happiness being dependent on some kind of success. U deserve to be happy by default and should reach for ur goals from a place of curiosity and love of urself and others instead of insecurity and fear of others.

Sometimes life will hit u with a hammer. When it does take time to regain ur hope in humanity in any way u feel is right and take as much time as u need. When u get back up u will be a stronger version of urself and ready to fail even harder again.

If ur interested in becoming a researcher/developer, I would be grateful for ur collaboration as I cannot do this alone. To get started read how to contribute and then setup development environment for Mac and Linux.

How to contribute

1. e-mail quinnliu@vt.edu ur schedule of when ur free after 7pm-11pm EST any day of the week for a 5 min. conversation to make sure our goals align. Also include a phone number :)

2. For now we are using the Git workflow model here and the Google python style guide here to contribute to this repository effectively.

Setup development environment for Mac and Linux

1. Go to the top right of this page and hit the Fork button. Then clone ur forked repository locally.

2. Install Python 3.5.2 and pip. U have installed successfully when u can:

   $ python3 --version
   Python 3.5.2
   

3. Install PyCharm Community

4. Locally run: pip3 install virtualenv

   • then cd wAlnut
   • then virtualenv wAlnut_environment
   • then source wAlnut_environment/bin/activate
   • then pip3 install -r requirements.txt
   • then nosetests --exe -s where -s tells nose to show print statements in test files and --exe flag look for tests in python modules that are executable
   • optionally type deactivate to leave the virtual environment

What

Introduction

This repo is an object oriented Python implementation of the computational framework outlined in Dileep George's PhD thesis How the Brain Might Work: A Hierarchical and Temporal Model for Learning and Recognition

After we understood the principles of flight we were able to build planes that are much faster and can carry more weight than birds. After we understand the human brain's computation principles we will be able to create intelligent machines much more intelligent than humans with unimaginable capabilities. At first it will not have any emotions like u and I. This will be a good thing because it will force us to ask if it would be moral to create an super intelligent machine with emotions to help us. It may not necessarily be a good thing for the human race. For example, look at how humans treat intelligent pigs. We are more intelligent emotional machines compared to pigs and the future super intelligent emotional machines will be much more intelligent compared to us. However if the emotional intelligent machine understands Hunter's Principle and uses it to figure out what is true and what is not true then we think it would be okay and safe to create a super intelligent emotional machine.

Key Theories

1. Theory 1: 1 common learning algorithm in the neocortex of the brain

   • Supportive:

     • 1992 Paper here from Department of Brain and Cognitive Sciences at MIT.
       • Summary: A portion of wires of the optic nerve are routed to the auditory cortex in ferrets. The neurons of this primary auditory cortex(A1) with vision input did not work exactly like neurons in primary visual cortex(V1). Neurons in rewired A1 had larger receptive field sizes and other differences. However, there are also similarities including:
         • rewired A1 neurons showed orientation and direction selectivity.
         • similar proportions of simple, complex, and non oriented cells between rewired A1 and V1.
         • implies "significant implications for possible commonalities in intracortical processing circuits between sensory cortices".

     • 1988 Paper here from Laboratoire des Neuroscience de la Vision at Universite de Paris.

       • Summary: Wires of the optic nerve were routed permanently to the main thalamic somatosensory nucleus in hamsters. After the hamsters grew up the neurons in the somatosensory cortex were recorded. The "somatosensory neurons responded to visual stimulation of distinct receptive fields, and their response properties resembled, in several characteristic features, those of normal visual cortical neurons."
         • "the same functional categories of neurons occurred in similar proportions, and the neurons' selectivity for the orientation or direction of movement of visual stimuli was comparable" between normal hamsters and rewired hamsters.
         • "These results suggest that thalamic nuclei or cortical areas at corresponding levels in the visual and somatosensory pathways perform similar transformations on their inputs".

     • 2008 PhD thesis here by Dileep George from Stanford University

       • There is a idea in optimization algorithms called No Free Lunch theorem. No Free Lunch theorem "state[s] that any two optimization algorithms are equivalent when their performance is averaged across all possible problems".
       • "On the surface, the NFL theorem seems to create problems for the idea of a common cortical algorithm. How can one mechanism/algorithm do very well on tasks as different as vision, audition and language? The answer comes from the part of the NFL theorem that talks about the assumptions that need to be exploited ... If the cortex is good at learning a wide variety of tasks using a common mechanism, then there must be something common about these seemingly different tasks.".

   • Not supportive:

     • 2002 Paper here from Vision, Touch and Hearing Research Centre at The University of Queensland.

       • "recent studies have revealed substantial variation in pyramidal cell structure in different cortical areas".

     • 2008 PhD thesis here by Dileep George from Stanford University

       • "Until we understand and explain the computational reason behind a large majority of such variations, the common cortical algorithm will have to be considered as a working hypothesis".

   • Conclusion: If cortex X(arbitrary cortex) of the neocortex(contains visual cortex, auditory cortex, somatosensory cortex, and others.) can be given non-normal sensory input usually given to say cortex Y and then learn to process this new input similarly to how cortex X would process it, then we can hypothesize that there is a common learning/predicting algorithm in all cortices of the neocortex.

2. Theory 2: The common learning algorithm efficiently learns about the world by making a specific set of assumptions about the world. This is the inductive bias of the common learning algorithm.

   • Supportive:

     • Thought Experiment here by D. Scott Phoenix from Vicarious.com
       • "Imagine we are designing a system to distinguish faces from non-faces and we are just using 32 by 32 pixel images and 256 shades of grey." This is a very reduced version of what ur visual brain is doing. "Even with that reduction it is still 256^1024 different possible combinations you need to be able to distinguish between". Because 32 by 32 with 256 different values is equivalent to the permutation (choose value between 0 and 256)(0-256)...32*32 more times.
       • 256^1024 is approximately 10^2400 which is a very large number. On the other hand a human lives about 2.5 * 10^9 seconds (80 years) and has about 10^11 neurons. And in about 3.1 * 10^7 seconds (1 year) "we are able to solve this problem space and distinguish readily between faces and not faces. What this implies is that our brains are making a lot of assumptions."

   • Not Supportive:

   • Conclusion: We need to find all the assumptions the brain's common learning algorithm makes about the world by asking the question "What is the basic set of assumptions that are specific enough to make learning feasible in a reasonable amount of time while being general enough to be applicable to a large class of problems?" ~ Dileep George

3. Theory 2 Assumption 1: A common learning algorithm must be trained on unlabeled movies of the world it is expected to learn to predict and classify.

   • Supportive:
     • A baby has to learn this way for the first year of life before it understands speech.
     • 2008 PhD thesis here by Dileep George from Stanford University
       • Figure 1:
       • Notice how the scaling of "A" in (c) would be the exact same input as (b) if u translate the input to the left by 1 pixel.
   • Not supportive:
   • Conclusion: Yes, specific enough to make learning feasible and general enough for large class of problems.

4. Theory 2 Assumption 2:
   Manifold = all the images generated by the same object in a high-dimensional space.

   If object A occurs close together to object B in time than object C then object A and object B are more similar. This information is used to form manifolds aka invariant representations of the objects.

   • Supportive:
     • A baby has to learn this way for the first year of life before it understands speech.
   • Not supportive:
   • Conclusion: Yes, specific enough to make learning feasible and general enough for large class of problems.

5. Theory 2 Assumption 3: If pattern B follows pattern A in time, then they are causally related and in the future pattern A should predict pattern B.

   • Supportive:
     • The brain is constantly making predictions about the future.
   • Not supportive:
     • How does the brain make predictions multiple time steps into the future?
   • Conclusion: Yes, specific enough to make learning feasible and general enough for large class of problems.

6. Theory 2 Assumption 4: Complex invariant representations are made up of less complex invariant representations in a hierarchy.

   • Supportive:
     • More ideas in the universe than neurons in the brain. Also the universe has natural hierarchies.
     • The neocortex has a hierarchal structure.
   • Not supportive:
   • Conclusion: The common learning algorithm first learns invariant representations of object components then learns invariant representations of more complex objects in terms of the invariant representations of the components.

7. Theory 2 Assumption 5: During the current time step of the common learning algorithm, it will always receive input about what muscles were used in the last time step.

   • Supportive:
     • Consciousness doesn't get confused when the eye is moving around. This means every region in the neocortex must be receiving input about the eye's sensori-motor movement so it can use it to accurately predict the future.

• Not supportive:
• Conclusion: This assumption is specific enough to make learning feasible in a reasonable amount of time while being general enough to be applicable to a large class of problems.

8. Theory 2 Assumption 6: Every input pattern is encoded into a sparse distributed representation of the input stimuli.

   • Supportive:
     • Not possible to process every pixel in the input image.
     • Sparse distributed representations can represent vision data, audio data, touch data and in general any time of data can be encoded into a sparse distributed representation.
   • Not supportive:
   • Conclusion: Specific enough to make learning feasible in a reasonable amount of time while being general enough to be applicable to a large class of problems.

9. Theory 2 Assumption 7: Consciousness is the illusion of an object having free will.

   • Supportive:
   • Not supportive:

Code Outline

If ur confused what a file/folder is for:

- walnut/
  - datasets/ = contains zipped datasets generated by https://github.com/WalnutiQ/time_series_data
  - images/explanatory/ = contains explanatory images about how the model looks
  - model/ = object oriented model implementing the key theories
  - tests/ = unit tests for code in model folder
    - experiments/ = using the OO model for algorithm experiments
  - utils/
- wAlnut_environment/ = folder only on ur local repo generated by virtualenv
                        containing libraries we are using
- .gitignore = contains names of files/folders not to add to this repository
               but keep in ur local wAlnut folder
- .travis.yml = for running this code on Travis's servers to make sure all tests pass
- LICENSE.txt = GNU General Public License version 3
- README.md = the file ur reading right now
- requirements.txt = list of python library versions we are using. These are installed
                     in ur virtual environment during setup