See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/337023590

How Artificial Intelligence Works

Article · October 2019

CITATIONS READS
4 13,568

1 author:

Nathan-Ross Adams
University of Johannesburg
3 PUBLICATIONS 4 CITATIONS

SEE PROFILE

All content following this page was uploaded by Nathan-Ross Adams on 13 November 2019.

The user has requested enhancement of the downloaded file.

How Artificial Intelligence <Currently> Works [Working Paper]

For this paper, a basic introduction to how AI works is most apt. The writer follows
tradition by introducing AI by comparing it to a human. In particular, the writer draws
comparisons with the cognitive structure, cognitive processing organs, and learning
techniques of humans.

1 Artificial Neural Networks

The human brain comprises of a network of neurons that process information. Since the
1940s, AI scientists have been working on replicating the structure and the function of
the human brain to manufacture similar cognitive capabilities for AI.1 This AI research
subfield is known as 'Artificial Neural Networks' (ANNs). Initially, the development of
ANNs was slow because of weak computer processing power. Recently, however, AI
researchers have rapidly developed fully-functional ANNs because of advances in the
processing power of computers in the 4IR.2

Decombe illustrates the uses of neural networks:

Artificial Neural Networks are used for a variety of tasks, a popular use is for
classification. You can collect datasets of images for example of different breeds of dogs
and then train a neural network on the images, then if you supply a new image of a dog it
will give a statistical score on how closely the new image matches the model and then
will output what breed of dog the image is.

Further, advanced ANNs are capable of retaining short memories.3

1
Dacombe J ‘An introduction to Artificial Neural Networks (with example)’ ​Medium​ 23 October 2017
available at
https://medium.com/@jamesdacombe/an-introduction-to-artificial-neural-networks-with-example-ad459bb
6941b​ (accessed 2 September 2019).
2
Dacombe J ‘An introduction to Artificial Neural Networks (with example)’ ​Medium​ 23 October 2017
available at
https://medium.com/@jamesdacombe/an-introduction-to-artificial-neural-networks-with-example-ad459bb
6941b​ (accessed 2 September 2019).
3
ANNs capable of memory are referred to as ‘Recurrent Neural Networks’ (RNNs). For a brief
introduction to RNNs, see generally Banerjee S ‘An Introduction to Recurrent Neural Networks’ ​Medium
23 May 2018 available at
2 The Organs of Artificial Intelligence

The sensory organs of humans inspired the model of AI organs. However, note that AI
has the potential to exceed human capabilities when the processing power of AI
exceeds that of humans. Below I explore the organs of AI compared with human
capabilities.

First, humans can recognise that someone is speaking to them and identify who the
speaker is by the timbre of their voice. This recognition starts with the act of speaking,
then the acts of hearing, processing, and identification. In AI research, this is the
subfield of speech recognition, that is, the ability of a computer to recognise the
particular nuances of a human's voice.4

Speech recognition has a complex history of development. However, due to the

advances in technology and the increase in access to data, it has become increasingly
intelligent. The work of South African speech recognition experts speaks to the
advancement of speech recognition technology. These experts have developed speech
recognition models that can understand the nuances of South African English accents
for transcription of speech.5

Speech recognition works similar to how a child would identify their parents and learn a
language. The child would initially hear words spoken by their parents, absorb
information about their timbre, then his/her brain would form patterns and connections. 6
Similarly, speech recognition technology would 'hear' by using microphones for

https://medium.com/explore-artificial-intelligence/an-introduction-to-recurrent-neural-networks-72c97bf091
2​ (accessed 1 August 2019).
4
Lynn J ‘Can We Talk: Speech Recognition Technology is Changing Your Relationship with Your
Computer’ (1999) 14 ​Commercial Law Bulletin​ 14-7.
5
Kamper H & Niesler TR ‘The Impact of Accent Identification Errors on Speech Recognition of South
African English’ 110 ​South African Journal of Science 6​ 3-9.
6
Globalme ‘Speech Recognition Technology Overview’ available at
https://www.globalme.net/blog/the-present-future-of-speech-recognition​ (accessed 10 September 2019).
real-time recognition or recordings of speech for later recognition. After that, this
technology trains itself to identify and distinguish between different voices by forming
patterns and connections.

An example of speech recognition technology is the Google Assistant that can be used
to unlock a mobile phone7 and autonomously create restaurant reservations8. Another
example is the voice-to-text transcription software, Otter AI. Otter AI can transcribe
recordings of conversations (e.g. interviews, lecturers, meetings) and identify then
separate each speaker according to the timbre of his/her voice.9

Secondly, computers may excel at learning from number-filled spreadsheets of data;

however, humans generally communicate with words, not numbers.10 The comparative
subfield of AI research is called 'Natural Language Processing' (NLP). NLP enables
computers to understand and communicate in human language. NLP realises this goal
through programming computers to understand and apply rules of syntax (grammar),
semantics (the meaning of words), and pragmatics (context and subtext).11 An example
of an NLP technology includes Grammarly, an AI-powered writing assistant that

7
‘The Google Assistant is a virtual assistant powered by artificial intelligence and developed by Google
that is primarily available on mobile and smart home devices. Unlike Google Now, the Google Assistant
can engage in two-way conversations.’ Oloo V ‘How to unlock your phone with your voice using Google
Assistant’ ​Dignited​ 8 September 2018 available at
https://www.dignited.com/34613/unlock-smartphone-voice-google-assistant/​ (accessed 7 June 2019).
8
Lumb D ‘Google Assistant Can Now Make Vocal Restaurant Reservations in 43 States’ ​Techradar​ 6
March 2019 available
https://www.techradar.com/news/google-assistant-can-now-make-restaurant-reservations-in-43-states
(accessed 10 September 2019).
9
Su J ‘CEO Tech Talk: How Otter.ai Uses Artificial Intelligence To Automatically Transcribe Speech To
Text’ ​Forbes 1​ 8 June 2019 available at
https://www.forbes.com/sites/jeanbaptiste/2019/06/19/ceo-tech-talk-how-otter-ai-uses-artificial-intelligence
-to-automatically-transcribe-speech-to-text/#76d8d1c38729​ (accessed 31 July 2019).
10
Raval S ‘Natural Language Processing’ ​YouTube​ 26 March 2019 available at
https://www.youtube.com/watch?v=bDxFvr1gpSU​ (accessed 3 September 2019).
11
Gour R ‘What is Natural Language Processing in Artificial Intelligence?’ ​Medium 1 ​ 9 March 2019
available at
https://medium.com/@rinu.gour123/what-is-natural-language-processing-in-artificial-intelligence-b13dc4a
a1c81​ (accessed 3 September 2019).
automatically detects grammar, spelling, punctuation, word choice, and style mistakes.12
Grammarly's algorithms flag issues in the written text and suggest corrections based on
context and selected writing styles.13

Finally, humans can generally see with their eyes and process what they see.
'Computer Vision' is a subfield of AI that focuses on enabling computers to interpret and
understand the visual world.14 The visual world consists of image data which includes
images, videos (a series of image frames), and three-dimension objects.15 This data can
be fed and processed in real-time by cameras or processed after the fact by image data
uploaded into the computer vision system. Thus, '[a]t an abstract level, the goal of
computer vision problems is to use the observed image data to infer something about
the world'.16

An example of Computer Vision technology is Facebook's facial recognition software.

This software can analyse the details of people's faces in uploaded photos such as the
distance between their eyes, nose, and other facial features.17 In ​Patel v Facebook,​ a
recent case dealing with whether Facebook users can sue Facebook for consent issues
related to facial recognition software, the USA Ninth Circuit Court of Appeals articulated
that ‘[o]nce a face template of an individual is created, Facebook can use it to identify

12
Marr B ‘The Amazing Ways Google And Grammarly Use Artificial Intelligence To Improve Your Writing’
Forbes​ 12 November 2018 available at
https://www.forbes.com/sites/bernardmarr/2018/11/12/the-amazing-ways-google-and-grammarly-use-artifi
cial-intelligence-to-improve-our-writing/#46c98ded3bb0​ (accessed 3 September 2019).
13
Hill S ‘How Grammarly & Google are Using Artificial Intelligence for Flawless Writing’ ​Big Data Made
Simple ​13 December 2018 available at
https://bigdata-madesimple.com/how-grammarly-google-are-using-artificial-intelligence-for-flawless-writin
g/​ (accessed 3 September 2019).
14
Raval S ‘Learn Computer Vision’ ​YouTube​ 14 July 2019 available at
https://www.youtube.com/watch?v=FSe_02FpJas​ (accessed 3 September 2019).
15
Raval S ‘Learn Computer Vision’ ​YouTube​ 14 July 2019 available at
https://www.youtube.com/watch?v=FSe_02FpJas​ (accessed 3 September 2019).
16
Brownlee J ‘A Gentle Introduction to Computer Vision’ ​Machine Learning Mastery​ 5 July 2019 available
at ​https://machinelearningmastery.com/what-is-computer-vision/​ (accessed 5 September 2019).
17
Ingber S ‘Users Can Sue Facebook Over Facial Recognition Software, Court Rules’ ​NPR ​8 August
2019 available at
https://www.npr.org/2019/08/08/749474600/users-can-sue-facebook-over-facial-recognition-software-cour
t-rules​ (accessed 4 September 2019).
that individual in any of the other hundreds of millions of photos uploaded to Facebook
each day, as well as determine when the individual was present at a specific location.’18

3 The Humanoid Representation of Artificial Intelligence

Computer scientists can present AI in human form by using robot technology. These
AI-powered robots are known as smart robots.19 Historically, AI and robotics were
viewed as one field, as can be seen in Part 2.1; however, of late, they are considered as
two separate fields. Smart robots are thus the convergence of these fields. By
illustration, Sophia is a smart robot, presented in human form, that was recently granted
citizenship in Saudi Arabia.20

4 How Artificial Intelligence Learns

4.1 Deep Learning

As the ANNs described elsewhere become more sophisticated, they are programmed to
learn more challenging problems. This subfield of AI research is 'deep learning'. Deep
learning techniques aim to replicate the learning development pathways of humans with
a focus on the visual or symbolic work. For example, if an ANN scans images from left
to right then top to bottom, it is called a convolution neural network (CNN). A CNN is
used to recognise objects using computer vision.

4.2 Machine Learning

18
​Patel v Facebook Inc​ No. 18-15982 (9th Cir. 2019) p17.
19
​ ed (1985) Springer, USA 1-30.
Hunt V ​Smart Robots: A Handbook of Intelligent Robotic Systems 1
20
Stone Z ‘Everything You Need to Know about Sophia, The World's First Robot Citizen’ ​Forbes​ 7
November 2017 available at
https://www.forbes.com/sites/zarastone/2017/11/07/everything-you-need-to-know-about-sophia-the-world
s-first-robot-citizen/ (accessed 2 January 2019).
View publication stats

As opposed the visual learning, AI can also learn from databases or data sets, that is,
machine learning. Machine learning requires large data sets (Big Data) to learn.
Machine learning empowers AI to learn from patterns in data, then make predictions
based on what it has learned. For example, if a human were to consider a graph of
advertising spend versus sales in a company, the human can plot the data then observe
the data in up to three dimensions. However, AI with machine learning observes the
data then creates patterns in hundreds of dimensions. Machine learning thus exceeds
human capabilities in this regard.