I like doing research on and writing about collective intelligence in human societies. I am even learning to program in Python in order to know how to code collective intelligence in the form of an artificial neural network. I decided to take on my own intelligence as an interesting diversion from the main course. I hope I can assume I am an intelligent structure. Whilst staying essentially coherent, i.e. whilst remembering who I am, I can experiment a bit with many different versions of myself. Of course, a substantial part of the existential Me works like a shuttle train, going back and forth on the rails of my habits. Still, I can learn heuristically on my own experience. Heuristic learning means that as I learn something, I gain new knowledge about how much more I can learn about and along the lines of the same something.
I want to put into a Python code the experience of heuristic, existential learning which I exemplified in the update titled ‘Something even more basic’. It starts with experience which happens through intentional action from my part. I define a vector of action, i.e. a vector of behavioural patterns, associated with the percentage of my total time they take. That percentage can be understood, alternatively, as the probability that any given minute in the day is devoted to that specific action. Some of those patterns are active, and some are dormant, with the possibility of being triggered into existence. Anyway, it is something like A = {a1, a2, …, an}. Now, in terms of coding in Python, is that vector of action a NumPy array, or is it a Pandas data frame? In terms of pure algorithmic technique, it is a trade-off between computational speed, with a NumPy array, and programmatic versatility in the case of a Pandas data frame. Here are a few points of view expressed, as regards this specific matter, by people smarter than me:
>> https://www.geeksforgeeks.org/difference-between-pandas-vs-numpy/
>> https://towardsdatascience.com/performance-of-numpy-and-pandas-comparison-7b3e0bea69bb
>> https://vitalflux.com/pandas-dataframe-vs-numpy-array-what-to-use/
In terms of algorithmic theory, these are two different, cognitive phenomena. A NumPy array is a structured collection of numbers, whilst a Pandas data frame is a structure combining many types of data, e.g. string objects with numbers. How does it translate into my own experience? I think that, essentially, my action is a data frame. I take purposeful action to learn something when I have a logical frame to put it in, i.e. when I have words to label what I do. That leads me to starting at even more elementary a level, namely that of a dictionary as regards my actions.
Anyway, I create a notebook with JupyterLab, and I start like a hamster, with stuffing my cheeks with libraries:
>> import numpy as np
>> import pandas as pd
>> import os
>> import math
Then, I make a first dictionary:
>> Types_of_action=[‘Action 1′,’Action 2′,’Action 3′,’Action 4′,’Action 5’]
A part of my brain says, at this point: ‘Wait a minute, bro. Before you put labels on the things that you do, you need to be doing things. Humans label stuff that happens, essentially. Yes, of course, later on, me can make them metaphors and abstract concepts but, fundamentally, descriptive language comes after experience’. Well, dear part of my brain, this is a valid argument. Things unfold into a paradox, just as I like it. I need raw experience, primal to any logical structuring. How to express it in Python? I can go like:
>> Raw_experience=np.random.rand(np.random.randint(1)) #This is a NumPy array made of random decimal values, and the number of those values in the array is random as well.
I check. I type ‘Raw_experience’ and run it. Python answers:
>> array([], dtype=float64) # I have just made a paradox: a totally empty array of numbers, i.e. with no numbers in it, and yet those inexistent numbers have a type, namely that of ‘float64’.
I try something less raw and more cooked, like:
>> Raw_experience_50=np.random.rand(50) # I assume a priori there are 50 distinct units of raw experience
>> Raw_experience_50 # yields…
>> array([0.73209089, 0.94390333, 0.44267215, 0.92111994, 0.4098961 ,
0.22435079, 0.61447481, 0.21183481, 0.10223352, 0.04481922,
0.01418667, 0.65747087, 0.22180559, 0.6158434 , 0.82275393,
0.22446375, 0.31331992, 0.64459349, 0.90762324, 0.65626915,
0.41462473, 0.35278516, 0.13978946, 0.79563848, 0.41794509,
0.12931173, 0.37012284, 0.37117378, 0.30989358, 0.26912215,
0.7404481 , 0.61690128, 0.41023962, 0.9405769 , 0.86930885,
0.84279381, 0.91174751, 0.04715724, 0.35663278, 0.75116884,
0.78188546, 0.30712707, 0.00615981, 0.93404037, 0.82483854,
0.99342718, 0.74814767, 0.49888401, 0.93164796, 0.87413073])
This is a short lesson of empiricism. When I try to code raw, completely unstructured experience, I obtain an empty expanse. I return to that interesting conversation with a part of my brain. Dear part of my brain, you were right to point out that experience comes before language, and yet, without language, i.e. without any logical structuring of reality, I don’t know s**t about experience, and I cannot intelligibly describe it. I need to go for a compromise. I make that experience as raw as possible by making it happen at random, and, in the same time, I need to give it some frame, like the number of times those random things are supposed to happen to me.
I defined a dictionary with 5 types of action in it. Thus, I define a random path of happening as an array made of 5 categories (columns), and 50 rows of experience: Raw_experience_for_action=np.random.rand(50,5).
I acknowledge the cognitive loop I am in, made of raw experience and requiring some language to put order in all that stuff. I make a data frame:
>> Frame_of_action = pd.DataFrame (Raw_experience_for_action, columns = [Types_of_action]) # One remark is due, just in case. In the Python code, normally, there are no spaces. I put spaces, somehow in phase with interpunction, just to make some commands more readable.
I check with ‘Frame_of_action.info()’ and I get:
>> <class 'pandas.core.frame.DataFrame'>
RangeIndex: 50 entries, 0 to 49
Data columns (total 5 columns):
# Column Non-Null Count Dtype
--- ------ -------------- -----
0 (Action 1,) 50 non-null float64
1 (Action 2,) 50 non-null float64
2 (Action 3,) 50 non-null float64
3 (Action 4,) 50 non-null float64
4 (Action 5,) 50 non-null float64
dtypes: float64(5)
memory usage: 2.1 KB
Once I have that basic frame of action, what is my next step? I need to learn from that experience. The frame of action is supposed to give me knowledge. What is knowledge coming from action? That type of knowledge is called ‘outcomes’. My action brings an outcome, and I evaluate it. Now, in a baseline algorithm of artificial neural network, evaluation of outcomes happens by pitching them against a predefined benchmark, something like expected outcome. As I am doing my best to be an intelligent structure, there is that aspect too, of course. Yet, there is something else, which I want to deconstruct, understand, and reconstruct as Python code. There is discovery and exploration, thus something that I perceive as entirely new a piece of experience. I don’t have any benchmark I can consciously pitch that experience against.
I can perceive my fresh experiential knowledge in two different ways: as a new piece of data, or as an error, i.e. as deviation from the expected state of reality. Both mathematically, and algorithmically, it is a difference. Mathematically, any number, thus any piece of data, is the result of an operation. If I note down, in the algorithm of my heuristic learning, my new knowledge as literally new, anyway it needs to come from some kind of mathematical operation: addition, subtraction, multiplication, or division.
As I think about myself learning new stuff, there is a phase, in the beginning, when I have some outcomes, and yet I don’t have any precise idea what those outcomes are, exactly. This is something that happens in coincidence (I don’t even know, yet, if this is a functional correlation) with the actions I do.
As I think about all that stuff, I try to make a loop of learning between action and outcomes, and as I am doing it, I realize I got it all wrong. For the last few weeks, I have been assuming that an intelligent structure can and should be coded as a loop (see, for example, ‘Two loops, one inside the other’). Still, as I am trying to code the process of my own heuristic learning, I realize that an algorithmic loop has fundamental flaws in that respect. Essentially, each experimental round – where I pitch the outcomes of my actions against a pre-defined expected outcome – is a separate loop, as I have to feed forward the resulting error. With many experimental rounds, like thousands, making a separate loop for each of them is algorithmically clumsy. I know it even at my neophytic stage of advancement in Python.
When I don’t know what to do, I can ask around. I can ask people smarter than me. And so I ask:
>> https://hackernoon.com/building-a-feedforward-neural-network-from-scratch-in-python-d3526457156b
After rummaging a bit in the content available under those links, I realize that intelligent structures can be represented algorithmically as classes (https://docs.python.org/3/tutorial/classes.html ), and it is more functional a way than representing them as loops. From the second of the above-mentioned links, I took an example of algorithm, which I allow myself to reproduce below. Discussing this algorithm will help me wrapping my own mind around it and developing new understandings.
A neural network is a class, i.e. a type of object, which allows creating many different instances of itself. Inside the class, types of instances are defined, using selves: ‘self.input’, ‘self.output’ etc. Selves are composed into distinct functions, introduced with the command ‘def’. Among the three functions defined inside the class ‘NeuralNetwork’, one is particularly interesting, namely the ‘_init_’. As I rummage through online resources, it turns out that ‘_init_’ serves to create objects inside a class, and then to make selves of those objects.
I am trying to dissect the use of ‘_init_’ in this specific algorithm. It is introduced with three attributes: self, x, and y. I don’t quite get the corresponding logic. I am trying with something simpler: an algorithm I found at https://www.tutorialspoint.com/What-is-difference-between-self-and-init-methods-in-python-Class :
I think I start to understand. Inside the ‘_init_’ function, I need to signal there are different instances – selves – of the class I create. Then, I add the variables I intend to use. In other words, each specific self of the class ‘Rectangle’ has three dimensions: length, breadth, and unit cost.
I am trying to apply this logic to my initial problem, i.e. my own heuristic learning, with the bold assumption that I am an intelligent structure. I go:
>> class MyLearning:
def _init_(self, action, outcomes, new_knowledge = 0):
self.action = action
self.outcomes = outcomes
self.new_knowledge = new_knowledge
def learning(self):
return action-outcomes
When I run this code, there is no error message from Python, which is encouraging for a newbie such as I am. Mind you, I have truly vague an idea of what I have just coded. I know it is grammatically correct in Python.
