You are on page 1of 12

Running Head: Functions and Applications of Machine Learning

Functions and Applications of Machine Learning

By Caleb Anton
Functions and Applications of Machine Learning

What I Know, Assume, And Imagine

Ive always appreciated the power behind computation and how weve managed to create highly

automated systems to stimulate human development. However, I always assumed that automating these

processes required a great deal of understanding from the user or programmer. Little did I know, that

weve managed to automate the very process of learning itself, and we can use that to solve problems and

recognize patterns we cant even begin to understand. After watching an artificial neural network learn to

play and eventually speed run a Mario level, I became very intrigued with the functions and applications

of machine learning.

Before officially researching my topic, I had already picked up on a few areas that machine

learning is being used today, whether it be from news articles or YouTube videos. I also understood that

many machine learning systems were based on the human brain and simulate many of the biological

functions that allow us to adapt to new ideas, situations, or problems. However, I felt that I needed to

learn more about the topic to better appreciate its inner workings and how it has and will change the

world.

I formed many questions about my topic. How do machine learning algorithms function? What

are the different types of algorithms? How is it currently affecting the world, and how will it affect the

world in the future. Will there be any breakthroughs? Even though my questions focused on many

different aspects of machine learning, they all fit under one overarching question: What are the processes

involved with and applications of machine learning? After searching, I was able to come up with an

answer that I felt satisfied with.

The Story of My Search

I began doing some background research with a simple google search for machine learning. It

became quickly apparent that I didnt understand much of the jargon used in the scholarly articles that

came from an unrefined search. From there, I used different glossaries from college resources to come up
Functions and Applications of Machine Learning

with a better vocabulary to tackle articles targeted at people that were more experienced with machine

learning. In order to understand a lot of the vocabulary that was defined in the glossaries, I needed to

understand different aspects of machine learning. This lead me to my first question: How do machine

learning algorithms work?

After searching through many scholarly articles, I found that it would be much easier to

understand the math behind how machine learning algorithms work if described with the sole purpose of

educating. I settled on a lecture series for students taking a machine learning class at Stanford. However,

after making my way through about half of the lectures, I realized that the information that I was learning

would be too in depth for the broad scope of my topic. I decided to use the information I picked up from

the lectures to go back to the glossaries that were too complicated for me to understand at first. I used this

information and honed in on a generic, yet simple to understand type of algorithm called error

backpropagation. On top of that, I learned about many different types of machine learning algorithms and

how to practically recreate them. Although it would be too difficult to describe in my iSearch, I planned

on using this information to extend my research by doing something practical.

Now that I was getting to know machine learning better from a technical aspect, I decided to

contact someone for an interview. I found that there was a local expert working with deep learning at the

Nasa Langley Research Center, and I was very excited to get their expert opinions on machine learning.

However, after not receiving a response for over a week, I decided to continue my search without the

interview. I figured that it would be important to research the questions that I wouldve asked the expert,

and that I could use this as an opportunity to find research by experts that would know more about the

questions I would ask the local expert.

From there, my research became a lot easier. Rather than focusing on the technical aspects, I

started to focus on how machine learning will and has changed the world we live in. There were no longer

any hard concepts to try and understand, so the main challenge that I faced in this portion of the research
Functions and Applications of Machine Learning

was making sure that my somewhat opinionated research was from unbiased sources. When looking at the

future of machine learning, it was very important to find a source that was based off of fact and data

analysis rather than expert opinions. Luckily, the Global Mckinsey Institute did a very thorough and well

known study on the future of automation and machine learning in many different types of fields and work

in those fields. I was able to confirm the validity of the source, because they were very reputable and well

known. Also, many other scholarly articles I found for this aspect in my research referenced the work

done by the Mckinsey Institute.

Finally, after getting a better understanding on how machine learning could be used and is

currently being used, I wanted to see how machine learning might benefit from the exponential growth of

computers. However, I quickly learned about the prospect of quantum computing, and how it would

benefit the world of machine learning much greater than increasing the speed of regular computers. I used

a couple different sources to learn how quantum computing will benefit machine learning. They werent

very technical and scholarly articles, but I felt that it was important that they were clear to understand

because I did not have enough time to dive into the subject of quantum mechanics. The different sources

that I ended up using contained matching information, and I thought that the matching information helped

to validate the sources.

After completing my research and organizing my thoughts, I felt that I had adequate information

to answer my question I set out to research.

The Search Results

Much of machine learning is based off of naturally occurring phenomenon. A lot of the basic

ideas and structure in machine learning algorithms are based off of the human brain. Because of this, it's

important to understand how biological neurons function. They are connected in complex arrangements,

and use electric pulses and specific chemicals to transmit impulses from one neuron to the next; in other

words, electrical pulses from other biological neurons cause the transfer of substances called
Functions and Applications of Machine Learning

neurotransmitters, or chemical messengers, across a structure called a synapse, to the dendrites, or the

receiving end, of other neurons, called downstream neurons. The arrival of the neurotransmitters in the

dendrite of the downstream neuron increases the tendency of the downstream neuron to send an electrical

pulse, or fire. If enough dendrites in a neuron receive neurotransmitters in a short period of time, the

neuron will fire.

Many machine learning algorithms use nodes, which are simple models of biological neurons and

are used in neural networks to perform a small part of some overall computational problem. Many

connected nodes, or artificial neurons, form a neural network. These connections have weights, meaning

that each has a parameter that determines how much notice a neuron pays to the activation it receives

from another neuron. These weights correspond to synapses in a biological neuron, and they are the

parameters that you multiply by the input signals in a given connection. If the weight is positive, the

connection is considered excitatory, while if the weight is negative, the connection is called inhibitory.

One of the most common and simple ways the neural network can be set up, is in a feedforward network,

which is a kind of neural network where information only travels forward. Because of this, nodes can be

numbered, and each node has weighted connections only to nodes with higher numbers. In practice, the

nodes of most feedforward nets are partitioned into layers, or sets of nodes, and layers may be numbered

in such a way that nodes in each layer are connected only to nodes in next layer, or the layer with the next

higher number. Commonly, successive layers are totally interconnected, meaning that each node in the

earlier layer is connected to every node in the next layer. Nodes are classified by which layer they are in.

The first layer has no input connections, making it the input layer, and it consists of input units. The last

layer has no output connections, making it the output layer, and are made up of output units. Layers in

between the input and output layers are termed hidden layers, and consist of hidden units. When the

network is operating, the activation of non-input neurons are computed using each neuron's activation
Functions and Applications of Machine Learning

function, or the function that describes the output behavior of the node. Most algorithms start by

computing the weighted sum of the inputs, which is the sum of the inputs multiplied by the weight

associated with each connection. The total net input is usually transformed in some way, using what is

sometimes called a squashing function. The simplest squashing function is a step function: if the total net

input is less than zero, or more often, some threshold T, then the output of the neuron is zero, otherwise it

is a one. These networks can be trained using an error backpropagation algorithm. (Wilson, 2012) When a

specific result is expected from the output units, the network is run, or propagated, with specific values

assigned, or clamped, to the input units. The squared error function is is used to compare the values of

output units with the expected values. After each propagation, the weights are updated in a way that helps

reduce the total error. Error backpropagation is a very simple, yet common example of a neural network

machine learning algorithm; however, there are many other types of machine learning algorithms.

(Mazur, 2016)

In order to make referring to machine learning algorithms less difficult, there exists a couple

different taxonomies. One way you can classify them is by learning style, and all algorithms fall under

three categories: supervised, unsupervised, or semi supervised learning. In supervised learning, the input

data is called training data, and it has a known result or label. A training process is repeated until the

desired level of accuracy is reached. For example, you could feed spam emails and non-spam emails into

a supervised training algorithm until it reaches a desired level of accuracy. Then, the algorithm could be

used to predict whether emails are spam emails or not. Some examples of algorithms under this category

are Logistic Regression and Error Back Propagation. In unsupervised learning, the input data is not

labeled and it doesn't have a known result. It deduces structures present in the input data and can be used

to extract general rules, reduce redundancy, or organize data by similarity. Some good of examples of

problems that can be solved using unsupervised learning are clustering (grouping), reducing the number

of random variables (dimensionality reduction), or discovering relations between variables in large


Functions and Applications of Machine Learning

databases ( or association rule learning.) In contrast to the spam and non-spam email example,

unsupervised learning could be used to group emails together by type when the categories arent

specified. Examples of unsupervised learning algorithms include the Apriori algorithm and k-Means. In

semi-supervised learning, the input data is a mixture of labeled and unlabeled examples. There is a

desired prediction problem, but the model must learn the structures to organize the data as well as make

predictions. They are used in classification and regression. Example algorithms extend to other flexible

methods that make assumptions about how to model unlabeled data. (Brownlee, 2017) Machine learning

algorithms are also grouped by similarity or how they work. There are a vast amount of functionally

different algorithms, but some prominent examples are tree-based methods, neural network inspired

methods, and regression algorithms. (Ray, 2017)

Because of machine learnings dynamicity, it's being applied to new problems all the time. Many

of these problems seem easy for humans to do, like interpreting languages, but pose an immense

challenge to those attempting to structurally define all the rules involved in the process in order to

automate it. However, some are very difficult to solve by humans, like finding arbitrary patterns in large

amounts of interconnected data, like ad or video recommendation algorithms. There are many different

problems that machine learning excels at, and has been able to eliminate a lot of unnecessary human

labor.

Because most malware are just slightly varied versions of previous malware, machine learning is

being used to very quickly find ways to defend against these, increasing data security. On top of that,

machine learning is benefiting personal security. It's being already being used in airports to spot

anomalies that people are very likely to miss. The stock market is another very classical machine learning

problem. Lots of data is being fed through deep learning bots that interpret and make trades at incredibly

high speeds and in large volumes. Smart cars are made possible through machine learning by developing

driving patterns and recognizing new conditions that would be very hazardous. Language has always been
Functions and Applications of Machine Learning

very difficult to interpret by computers. However, using deep learning, many greats strides are being

made on this issue. In fact, it's very likely that you've read an article written almost exclusively by a

machine learning algorithm. (Marr, 2016)

Google has been a very notable pioneer when it comes to deep learning, which is an algorithm

that consists of very complex interconnected neural network algorithms. They have many hidden layers.

They currently use it to optimize search results, recommend and promote videos on YouTube, etc. Also,

they have created google deepmind, which was made to adapt to many different problems. They have also

used it to play pictionary, Space Invaders, and many other fun/interesting ideas that have posed problems

for computers in the past. (DeepMind: Research, 2017)

A set of thorough studies have been done by the Global Mckinsy Institute that predict what the

use of machine learning in different fields. Knowing that machine learnings greatest potential across

industries includes improving forecasting and predictive analytics, 120 different use cases were analyzed.

Below is heat map applied to the problem types and industries.


Functions and Applications of Machine Learning

(Chui, n.d.) They have also predicted which jobs are the most likely to be replaced using machine

learning. The major factor involved with whether machine learning is applicable to different fields is the

amount of data produced. For now, many jobs will most likely stay safe; however, as the technology

develops, more and more jobs will become much easier to automate. (Columbus, 2017)

Even though machine learning algorithms are becoming more optimized, their speed and

effectiveness is greatly hindered by the processing power of the computer. In order for processing power

to increase, transistors, which are the building blocks of all computers, must become smaller. Luckily,

weve become very good shrinking transistors, and there is a law, Morgans Law, that states that the

number of transistors we can fit on a given chip doubles every two years (which doubles the processing

power every two years.) However, when they get small enough, they will be able to pass over regions

they aren't supposed to through a process called quantum tunneling. (Particles on the quantum scale are

able to ignore initial decreases in energies if it leads to an increase in energy.) Because transistors are

getting closer to reaching their theoretical size limit, Morgan's law is starting to break down. This has lead

many machine learning algorithms to rely on a connected network of computers, rather than using a single

processor. Luckily, machine learning suits itself really well to cloud computing. If anything, the

slowdown of computer processing power advancement only increases the need for machine learning.

(Future of Computing, 2016)

When transistor's eventually reach their limit, most research and development will be in the field

of quantum computing. Using many of the weird processes that occur at the quantum level, quantum

computing will be much faster when you have to look through or analyze a lot of information or do a lot

of things at once, like factoring. It won't be any better at doing regular straight forward tasks though.

However, one of the most exciting parts about quantum computing is that it will increase the speed of

machine learning algorithms exponentially (according to the number of qubits.) Once that happens,

machine learning will be able to used for so much more than we could even imagine. (Palmer, 2014)
Functions and Applications of Machine Learning

Growth As A Researcher

I have grown as lot as a result of this research project. Because I was put into a situation where I

had to dive into a topic and conduct formal research, I came to learn how interesting it can be. When I

ended up in a state of flow, the task became much less arduous and even rewarding. I can use this new

found craving for information on other research topics, which will allow me to give a genuine and

wholehearted effort into learning about new topics.

I also have learned how to benefit from formal research journals, and to gradually develop

vocabulary to tackle complex articles. This has opened up a wide variety of resources that I would have

previously ignored.

After completing my research, Im sure that I will continue to investigate the field of machine

learning and continue to grow as a researcher.


Functions and Applications of Machine Learning

References

Columbus, L. (2017, January 01). McKinsey's 2016 Analytics Study Defines The Future Of Machine

Learning. Retrieved June 12, 2017, from

https://www.forbes.com/sites/louiscolumbus/2016/12/18/mckinseys-2016-analytics-study-defines-t

he-future-machine-learning/#ce1c38414eb2

How quantum computing could change everything. (2014, October 28). Retrieved June 12, 2017, from

http://theweek.com/articles/443104/how-quantum-computing-could-change-everything

Marr, B. (2016, September 30). AI And Machine Learning Use Cases. Retrieved June 12, 2017, from

https://www.forbes.com/sites/bernardmarr/2016/09/30/what-are-the-top-10-use-cases-for-machine-

learning-and-ai/#400bc59a94c9

M. (2016, December 29). A Step by Step Backpropagation Example. Retrieved June 12, 2017, from

https://mattmazur.com/2015/03/17/a-step-by-step-backpropagation-example/

Michael Chui, James Manyika, and Mehdi Miremadi. (n.d.). Where machines could replace humans--and

where they can't (yet). Retrieved June 12, 2017, from

http://www.mckinsey.com/business-functions/digital-mckinsey/our-insights/where-machines-could

-replace-humans-and-where-they-cant-yet

Ray, S., N., Blog, G., & Gupta, D. (2017, May 04). Essentials of Machine Learning Algorithms (with

Python and R Codes). Retrieved June 12, 2017, from

https://www.analyticsvidhya.com/blog/2015/08/common-machine-learning-algorithms/

Research. (n.d.). Retrieved June 12, 2017, from https://deepmind.com/research/

A Tour of Machine Learning Algorithms. (2017, February 21). Retrieved June 12, 2017, from

http://machinelearningmastery.com/a-tour-of-machine-learning-algorithms/
Functions and Applications of Machine Learning

The future of computing. (2016, March 12). Retrieved June 12, 2017, from

http://www.economist.com/news/leaders/21694528-era-predictable-improvement-computer-hardw

are-ending-what-comes-next-future

You might also like