ABSTRACTION

WEEK4
• One of the most powerful problem-solving tools of computer science
is abstraction. Abstraction isn't about solving a particular problem faster
or with fewer resources.

• Instead, the goal of abstraction is to allow us to arrange information

more quickly and reliably in our heads and ignore irrelevant details.

• The purpose of abstraction is largely for helping humans think, rather

than helping computers work.
• We'll look at abstraction with the help of Jing, the
mayor of a small city.
• A city government is made of dozens, hundreds, or
even thousands of people who have their own roles,
who help the city in their own ways, and whose
responsibilities may overlap with one another.
• In a very small town, Mayor Jing might be able to keep track
of a dozen or so people's overlapping job descriptions, with
everyone reporting directly to the mayor.
• In a city with hundreds or thousands of employees, this
approach will certainly run into problems. What could go
wrong?
• Correct answer: Both of these.
• All of the downsides mentioned are potential
problems. With more and more employees, Mayor
Jing will need to start spending an increasing portion
of her time “micromanaging” her many employees.
• Overlapping responsibilities mean that every task
potentially needs to be coordinated with a different
set of employees. This confusion can sometimes make
it hard to ensure that all the important parts of a city's
business, like fixing water pipes, are ultimately
covered.
• As organizations grow, it's usual to break people up
into separate groups or departments, which each have
their own leaders, goals, and objectives. In the image
below, Mayor Jing has separated her employees into
three departments: Fire, Parks and Rec, and
Sanitation.
• Creating a box that everyone agrees on and labeling it “The Fire Department” is a
form of abstraction. Mayor Jing can tell the “Fire Department” to do something—like
focusing on rescuing cats in trees—without necessarily understanding all the details
of how that affects the individual humans, buildings, and fire trucks.
• This kind of change is not without drawbacks. What could go wrong with the
particular divisions shown here?
• Correct answer: Sewers might get inspected too
much, or not enough.
• In this arrangement, tree planting is the clear
responsibility of the Parks and Rec, the Fire
Department is in charge of fighting fires, and
Sanitation is fully in charge of fixing water pipes.
• Because sewer inspection ended up partially under
Water and Sanitation and partially under Parks, the
most likely problem is confusion about responsibilities
between these two groups.
• Splitting up a complicated system, like a city government,
into pieces involves tradeoffs, because different splits may
have different advantages and disadvantages. But it's
possible to imagine that some arrangements are clearly
better or worse than others, regardless of the tradeoffs.
• It's helpful to realize that many times the abstractions we
encounter are the result of historical accidents, other
people's personal preferences, or just because it's what
everyone else does. This is true in computer science and in
city governments. For example, it's easier for Mayor Jing
to follow the convention of having a Parks and Rec
Department and a Fire Department, just because those
are common departments that other cities and future
employees will be used to.
• It's important to organize complicated systems into
understandable parts with more or less well-defined tasks.
In both computer systems and human systems, these
abstractions exist to help humans understand the system,
despite its complexity.
• Both computer systems and human systems can end
up with lots of levels of abstraction: Mayor Jing
communicates with the head of the Fire Department,
who then communicates with the heads of her four
firehouses, who then each communicate with their
half-dozen or so shift leaders, who then communicate
with their shift employees.
 • What is a likely symptom of this proliferation of
abstraction layers in Mayor Jing's government?

Correct answer: It takes a long time to communicate with everyone.

If Mayor Jing can't communicate directly with everyone, and if all communication
always follows her chain of command, then communication will almost necessarily be
slower.
• The kinds of layers that Mayor Jing deals with in her organizational structure are
everywhere in computer science. When a company like Intel or AMD designs the
computer chip powering a computer or phone, that company also publishes an
abstraction layer: a set of commands for controlling the chip. These commands
are called assembly instructions or machine instructions.
• You can create a program by writing machine instructions directly, but this is way
too hard. It would be like Mayor Jing calling every employee to micromanage
every detail of rescuing a cat from a tree or repairing a water pipe.
• Instead of writing machine instructions, it's easier to write commands using a
programming language that's easier to write and read. A process
called compilation transforms the code you write into machine instructions.
Those machine instructions then direct the chip, as the Fire Department's chief
directs her subordinates.
• Computer systems and human systems do frequently have
many layers. Many languages that are designed to be
easier to use, like Python, don't ever turn the code you
write into machine instructions. Instead, there is a second
computer program, called an Interpreter, which is made
up of machine instructions.
• The program you write in Python is a bunch of Python
instructions that the Python interpreter understands. The
Python interpreter is a list of machine instructions that
the machine understands. This is like the multi-level
delegation where Mayor Jing and the firefighters don't
ever talk directly but instead go through the fire chief.
• There is a cost to having our computer code run on top of
Python, just as there is a cost to having Mayor Jing
communicate with her firefighters through the head of the
Fire Department. In both cases, the extra layer of
communication slows things down.
• If we write a program with Python, we may worry that our
program would run faster if we'd written it in some other
programming language. It's conceivable that our program
would run even faster if we'd painstakingly written the
machine instructions by hand instead of using a more
human-friendly programming language. Perhaps our
program would run faster still if we'd specifically designed
the hardware's structure to solve the problem rather than
relying on machine instructions!
• Each of these theoretically faster and more direct
approaches is also more difficult and complicated.
There's a resource tradeoff between the amount of
time it takes for humans to solve a problem and the
speed with which a computer can implement the
human-designed solution.
Making Decisions
• Computers can make decisions, and computers can
do things very very fast. Right now, a computer is
deciding what the solution to a mathematical equation
is. Somewhere else, a computer is deciding whether to
suspend someone's credit card to protect them from
fraud, and another computer is deciding whether an
image represents a stop sign or a bird.
• An important part of computer science is
understanding how computers can make the right
decisions, or at least pretty good ones.
• One of the ways computers (and sometimes humans)
make decisions is with a structure called a decision
tree. Decision trees encode a series of simple yes-or-
no questions that you can follow in order to answer a
more complex question. Here is a silly decision tree
that helps you decide which of eight different
creatures you're dealing with:
• If a computer were using this diagram while looking at
a creature that happened to be an adult blue whale,
the computer would start at the top box, the root of
the decision tree. The computer would ask whether
the creature was smaller than a bicycle. The answer is
definitely no, a blue whale is not smaller than a
bicycle.
• Next, the computer would follow the "no" path, the
arrow that goes to the right. The next box contains the
question "Does it swim?" The answer to that is
definitely "yes," blue whales swim. So the computer
would continue down the "yes" path, which is the
arrow that goes straight down. The computer then
correctly concludes that it's looking at a blue whale.
• A computer is using this decision tree. Which of the listed
creatures will require the computer to ask the most questions?
• A computer scientist can use this decision
tree to write a simple face recognizer. The
computer scientist first writes three
simpler tests to detect glasses, long hair,
and smiling. The decision tree organizes
these simple tests, allowing a computer to
distinguish between all the faces.
• The first question at the very top of the
decision tree is an especially important
one.
• Imagine that you want to make a decision tree to
distinguish the four faces above. If you want the root
of the decision tree to split the faces into two groups
of two, which of these questions should you ask?
• The first question in your decision tree will split the
faces into two distinct groups. What is a single
question that you could ask of both groups of two in
order to identify any single face?
• None of the choices you were given were able to further split
up both of the groups of two faces. You can nevertheless make
a decision tree that distinguishes all four faces, because the
tree can ask a different follow-up question based on the answer
to the first question. This flexibility makes decision trees quite
powerful.
• The shape of a decision tree can make a big difference in what happens when you
use it. If your computer program is using this decision tree, how many questions,
on average, will the computer program need to ask and answer in order to
distinguish a random face?
 How many of the faces take
more than three questions to
The decision tree from identify with this decision tree?
the previous page
distinguishes between
eight faces with exactly
three questions.
This decision tree
Distinguishes between
the exact same set of
eight faces; however, it
takes an average of
4 3/8​questions to
identify a random face
with this decision tree.
• This decision tree distinguishes six faces. Which face
will be chosen if the decision tree is used to identify
this different face:
• The decision tree on the previous page did not do a very
good job classifying this face. These faces are almost as
different as they can be! One reason that the decision tree
failed so badly was that the first few questions focused on
hair color and the decision tree contained no faces with
blond hair. The lack of blond-haired faces in the decision
tree made it easier to badly misclassify a blond-haired
face.
• This was a simple example, but if you pay attention to the
news, you'll see many real-world examples of exactly the
same kind of failure.
• It can be funny when computers make mistakes because they
were designed with limited information. For example, computer
programs designed to identify pictures sometimes "hallucinate"
sheep in every field. This is because the computer programs are
designed based on a bunch of pictures taken by people on
vacation. The pictures of fields taken by people on vacation mostly
also contain sheep.
• Other cases of computers failing to make good decisions are
more worrying. Many computer programs that do real-life facial
recognition don't work well for people with darker skin. This can
happen when the facial recognition program is designed around
a bunch of pictures the designers had easy access to, and those
pictures mostly contained white people.
• Decision trees are useful tools for computer scientists. They turn
simple yes-or-no decisions into more complex decisions involving
many different options. Computers are often better at answering
simple yes-or-no questions, so decision trees help computers
manage more complexity.
• Some computer scientists use decision trees designed by human
experts to help computers make smarter choices. Other computer
scientists use computer programs that do machine learning in
order to create the best decision trees for solving a new problem.
When you're using decision trees, the order of questions can
make a big difference in the number of questions you have to ask.
Class Task-Week 4
Decision trees aren't the right tool for every problem. For example,
how would you use a decision tree to sort a deck of cards?