Praise for Trading from Your Gut

Getting Your Brain to Listen to Your Gut
To better understand how to train intuition, let’s look at the
process used for developing neural networks. Neural networks sim-
ulate the connections between neurons in the brain and are a main-
stay of artificial intelligence in computer science. Neural networks
are excellent at categorizing and recognizing patterns, but, as in
humans, this is a learned skill.
To recognize patterns, neural networks first need to be trained.
This process involves submitting sample data to the neural network
representing the patterns that you want it to recognize. The training
C
HAPTER
2 • T
HE
P
URPOSE OF
G
UT
I
NTUITION
23
From the Library of Daniel Johnson

---

ptg
rewires the neural network to reflect the knowledge required to
understand the pattern presented. 
What do I mean by knowledge in this specific case? If you think
about the process your brain uses to determine if a piece of furni-
ture is a chair or a stool, you can get some idea of how your brain
uses knowledge. Your knowledge about chairs and stools is the infor-
mation you use to determine whether some new piece of furniture is
a chair, a stool, or neither. Specifically, this knowledge includes the
traits the two types of furniture hold in common: that both a chair
and a stool are something you sit on. The knowledge also includes
the concepts that differentiate a chair from a stool: that a chair is of
a specific height to fit a typical table and generally has a back,
whereas a stool has no back and is generally either shorter than a
chair and used for stepping to reach high objects, or taller than a
chair and used for sitting at a bar.
In a neural network, the knowledge consists of idealized models
that are categorized according to domain-specific categorization
schemes (or taxonomies), as well as the relationships among and
between those models. The picture that springs to mind when you
think of the word chair is an example of an idealized model. What
you use to differentiate a chair from a stool is an example of your
brain’s knowledge of the relationship between the idealized chair
and stool.
The neural network “learns” by being exposed to new examples
and the values for their hierarchical place within their particular cat-
egories. For example, you might feed the neural network a picture
of ten chairs and tell it that these are chairs, ten stools and tell it that
these are stools, and ten tables and tell it that these are tables. This
process enables the neural network to build internal models of what
24
T
RADING FROM
Y
OUR
G
UT
From the Library of Daniel Johnson

---

ptg
these examples represent. After the training, the network has inter-
nal models that represent its knowledge of the differences among
chairs, stools, and tables.
Taxonomies are categorization schemes. If you were building a
character-recognition neural network, one set of categories would
be the letters individually, as well as the set of uppercase letters and
lowercase letters. A given example might be categorized as a letter A
and also separately as an uppercased letter. Another example might
be categorized as a letter c and a lowercase letter. A perfect neural
network contains enough knowledge of the requirements for each
category to be able to determine whether a new sample fits the cat-
egory. The knowledge required to determine membership in a given
category is known as a model.
Much as you don’t generally picture a specific chair in your
mind when you think of the word chair, the models that a neural
network builds don’t represent specific letters that you have
encountered. Instead, they represent the idealized forms for each
letter. The model for the English letter S is the idealized curved
shape of the letter; it is similar to how you envision the letter S in
your mind. For a neural network to recognize an S in an arbitrary
typeface, the model it contains must be sufficiently complex so that
it can distinguish between an S and the other letters. The model
can’t be too specific, or it won’t be able to distinguish among an