mirror of
https://github.com/Clortox/tylerperkins.xyz.git
synced 2025-01-09 11:27:58 +00:00
Add modeling thought
This commit is contained in:
parent
821d460e9a
commit
36d56abc70
73
content/posts/thought-embedding-space/index.md
Normal file
73
content/posts/thought-embedding-space/index.md
Normal file
@ -0,0 +1,73 @@
|
||||
---
|
||||
title: "Modeling Thought"
|
||||
date: 2025-01-04T23:24:13Z
|
||||
draft: false
|
||||
---
|
||||
|
||||
I have played with the concept of how to model my own thought for a long time.
|
||||
Being that thought is a pretty complex yet predictable process, I have thought about
|
||||
what analogies or models I can use to better understand my own way of thinking, and
|
||||
why I do certain things.
|
||||
|
||||
My working model I have developed thus far (based on nothing but vibes) is
|
||||
linear algebra based. I believe this comes from both my CS/Math background causing
|
||||
a bias, as well as a bias from learning about how transformers model internal state.
|
||||
|
||||
So in this quick writeup, I would like to try and give a quick introduction to
|
||||
the concept I have, and over time I will develop it with some of the derivations
|
||||
I have found that come as a consequence.
|
||||
|
||||
#### Defining Thought
|
||||
|
||||
To begin, let me define what I mean when I say a *thought* or an *idea*.
|
||||
|
||||
> Thought
|
||||
>
|
||||
> Any single concept or abstract idea at a single point in time that can be
|
||||
> held inside of one's mind.
|
||||
>
|
||||
> Also see *Atomic Thoughts* as used in zettlekastens
|
||||
|
||||
Cool. What does this actually mean?
|
||||
|
||||
Your mind is almost always experiencing a thought,
|
||||
or sometimes it may feel as though there are several thoughts at once.
|
||||
You can think of this as a function of time and some other external stimuli.
|
||||
Lets call
|
||||
this function \(M\). Let \(t\) be time and \(e\) be some other
|
||||
state ocurring at this time.
|
||||
|
||||
$$
|
||||
M(t, e)
|
||||
$$
|
||||
|
||||
When I say a *thought* or *idea* I mean a single output from \(M\).
|
||||
|
||||
#### Modeling an Individual Thought
|
||||
|
||||
A single thought is a very complex thing! It has some general idea of "closeness"
|
||||
to other thoughts, as well as different "directions" that it can move in. These
|
||||
directions are things like how much of a certain adjative a thought containts,
|
||||
for example sentiment or much more complex adjatives like how much a thought
|
||||
relates to World War 2. These we can think of as "dimensions".
|
||||
|
||||
This is basically how embedding spaces work in language modeling.
|
||||
They are modeled as vectors, and I feel that this is a very natural analog to apply
|
||||
to thoughts as well.
|
||||
|
||||
We leave the paramaters of these vectors pretty abstract for now. They are \(d\)
|
||||
dimenions (presumably \(d\) is rather large) and each thought vector \(\boldsymbol{T}\) is created as a result of some paramaters to \(M\).
|
||||
|
||||
$$
|
||||
\forall t \ \forall e \ \exists \ \boldsymbol{T} \rightarrow \boldsymbol{T} = M(t, e)
|
||||
$$
|
||||
|
||||
Also note that \(\boldsymbol{e}\) is itself a vector.
|
||||
|
||||
Thoughts can be checked how similar they are by checking the dot product, which
|
||||
encodes the "closeness" we percieve in thoughts. Thoughts An amount of suprise
|
||||
can be found by how small the dot product is between the current thought,
|
||||
and the next thought.
|
||||
|
||||
Next, we will investigate how this model impacts modeling creativity.
|
||||
|
@ -5,6 +5,9 @@
|
||||
displayMath: [['\\[', '\\]'], ['$$', '$$']], // block
|
||||
inlineMath: [['\\(', '\\)']] // inline
|
||||
},
|
||||
macros: {
|
||||
vect: ["\\boldsymbol{#1}", 1]
|
||||
},
|
||||
chtml: {
|
||||
scale: 1.2
|
||||
}
|
||||
|
Loading…
Reference in New Issue
Block a user