Please enable JavaScript.

Coggle requires JavaScript to display documents.

SPERONI - Coggle Diagram

- - - - SQL Databases are exceptionally good at handling structured data and complex queries, making it easier to look for specific info across multiple tables.
    - - One key advantage it's the easier scalability across machines: if you need yours to grow, it will do it without a dramatic restructure.
- - - - state as an input
        
        #1 neural network
        value neural network
        
        output
        
        value
      - state as an input
        
        #2 neural network
        policy neural netowrk
        
        output
        
        policy
        (for each result the probability of doing it)
      - These neural networks
        should then fight against itself finding the best moves. After that the process is repeat again after backpropagation to the new generations.
- - - - EXAMPLES
        
        Wikipedia
        
        What technical detail permits Wikipedia to work?
        
        By adding an history of modifications on pages, a group of people can work together. It stops trolls from just deleting the work of hundreds of others.
        
        Russia e-democracy test in 2010
        Question of the test was: How should the law that govern policeman change?
        
        Too many contributors, so too many proposals.
        The problem arises: who can present proposals?
- - - - WHY? Because of an exponential growth in place
        
        Linear growth grows of the same quantity every day (month / year)
        
        Exponential growth grows of the same ratio every day (month / year).
        An expontential growth doubles (halves) every fixed time. This time is called the doubling (halving of the time).
        :warning:Yt1= 2 x Yt0 (t1-t0 = Doubling time)
        The Rule of 72
        To move from growth rate to the doubling (halving) of time you apply the rule of 72:
        
        If Y grows (decreases) by p% a day (mm/yyyy) it will double (halves) in 72/p days (mm/yyyy)
- - - - The Sigmoid Function
        Once we sum all the values multiplied by their
        weight, the value can be between (-inf, +inf). So
        we need to squeeze it into [0,1]
      - Back-propagation
        ● Neural networks are trained using a technique called Back-propagation.
        ● You need to have a set of input and output.
        ● Then you show the input, and see what output you receive. You compare (i.e. subtract) the output you get from the one you should get.
        ○ This is a measure of the error the network is doing.
        ○ As we change the input, we get a series of output and errors.
        ○ We average all the errors, and get an evaluation.
        ● Imagine that each set of parameters will get an evaluation.
        ● We need a strategy to change the parameters to lower this evaluation function until we find a minimum.
        ● Take an example:
        ○ Modify the weights
        ■ Get for each weight the error.
        ■ Average on “all” the various examples
        ○ Find what would be the optimal values for the previous layer
        ■ Modify the weight of the previous layer
        ■ Find what would be the optimal values for the
        previous layer
        ● Modify the weight of the previous layer
        ● Find what would be the optimal values for
        the previous layer
        ○ …
        
        In short:
        TEST IT> EVALUATE IT (comparing the output you want to get) > CHANGE THE PARAMETERS TO LOWER ERROR > Repeat
    - - After GOFAI
        
        The idea that Intelligence disembodied did not exist in nature
        
        This started a whole line of research in “embodied intelligence”
        
        After all a lot of the solutions to problems are inside the body itself
        
        Eventually many researcher realised that the only form of embodied intelligence were alive. This started the whole like of research in “Artificial Life” (a.k.a. Alife).
        
        Eventually many researcher realised that the only form of living system able to evolve were coming from chemical systems. This started the whole line of research in “Artificial Chemistries
      - Automatic translation
        word <-> word
        multiple w. <-> multiple w.
        (expression <-> expression)
        Translated for the computer into:
        Vector <-> Vector
        World and phrase translation with word2vec
        
        If we associate to each word a vector
        Vector size = vocabulary size
        If we assume the position has a meaning then...
        ● We can find the position of a word in space.
        ● Given a position we can find what words are around there.
        ● Given a word we can find the synonyms.
        ● We can sum words. (?)
        ● We can subtract words. (?)
- - - - Image Net Challenge 2011 (error 25%)
        
        Image Net Challenge 2012 (error 15%)
        
        Input data> Conv1,2,3,4,5> FC6>FC7>FC8
        
        Image Net Challenge 2013 (error 11%)
        
        Image Net Challenge 2014 (error 6%)
        
        22 layers
    - - AlphaGo Zero
        educated by only playing against itself