A Primer in Artificial Intelligence

Author: Zani Sharifi

What is Artificial Intelligence?

AI has been playing an increasingly important role for years. But there’s a lot of myths and misunderstandings surrounding the term. That’s why we decided to publish a short and easy to understand Primer in AI. We explain the term and show why it is probably the most important technology of this era. In this first part, we will answer the five most frequently asked questions.

We will start off with explaining the basic terminology – what do AI, Machine Learning (ML) and Deep Learning (DL) actually mean? Though the terms are closely related they refer to different areas of research.

AI is a broad field of research aiming to build intelligent machines that solve complex problems. In contrast to ML and DL, Artificial intelligence refers to the output of the problem-solving process. It does not explain HOW a problem is solved.

Machine learning encompasses numerous techniques used in AI applications. While AI’s goal is to imitate human intelligence, machine learning deals with re-creating the LEARNING PROCESS of humans. Different approaches of learning have emerged.

One such approach uses artificial neural networks (ANN). ANNs attempt to imitate the structure and functionality of the human brain, which is built on neurons connecting via synapses. Deep learning is a subset of machine learning enabling the search of solutions to difficult problems – through complex ANNs, i.e. neural networks with more neurons, layers, and connections.

Q1

Let us dive deeper into machine learning. Generally, there are three categories of machine learning algorithms: supervised, unsupervised and reinforcement learning.

Supervised learning uses labeled data, i.e. for each input a correspondent output exists. It is used to build predictive models; to derive relationships between inputs and outputs.

In unsupervised learning, there is only (unlabeled) input data. The aim is to uncover patterns. One unsupervised learning method is called clustering (finding groups of inputs with similar characteristics). Clustering can be used to segment customers. Another method, learning associations, can be used to find associations between data points. Applications include providing book recommendations based on previous purchasing behavior.

Reinforcement learning allows algorithms to learn by trial and error. The system receives rewards for performing a task correctly (based on predefined measures) and is punished for bad performance. Reinforcement learning is applicable if the output of a system is supposed to be a sequence of actions and is often used in gaming.

Q2

Time to talk about the three levels of AI:

The first level of AI is referred to as artificial narrow intelligence (ANI), or weak AI. ANI applications specialize in one single area such as playing chess, driving autonomously or holding conversations. Narrow AI is only applicable within a limited context. All of today’s AI applications belong to this category.

Human-level AI is a step ahead and referred to as artificial general intelligence (AGI), or strong AI. A computer possessing Artificial General Intelligence can perform human-level assignments across different disciplines without human aid. There is a debate among researchers on whether general AI is possible or not.

Nick Bostrom, an AI expert, has coined the term artificial superintelligence (ASI). Superintelligence is defined as an intellect, which is smarter than all human beings in all areas including creativity and social skills. For the moment being, both AGI and ASI remain theoretical constructs without reliable proof for their possible existence in the future.

Q3

Since we now know understand the different terms around AI, let us talk about the capabilities of AI applications. What skills does a machine or application need to possess to be ‘intelligent’?

Natural Language Processing (NLP)
Creating programs, which are capable of understanding and interacting in human language is one of the long-standing goals of AI. NLP focuses on interactions between human or natural languages and computers and is particularly important in the programming of computers so that they can effectively process large data that is presented in natural languages

Knowledge representation and Automated Reasoning
Simply put knowledge representation means that a program stores what it knows and hears. Reasoning is defined as the ability to make inferences. Building on knowledge representation, automated reasoning means that the process is automated, i.e. a program uses the stored information to answer questions and to come up with conclusions on its own.

Computer vision
Computer vision is the skill to acquire, process, analyze and understand digital images as well as to extract high-dimensional real-world data. The field can be dated back to the 1960s. While the process’ theoretical aspects were understood, the lack of computing power hindered good results.

Robotics
Robots are physical systems, which can assert physical force to manipulate objects with the use of so-called effectors (e.g. hands or legs). Further, robots are equipped with sensors allowing them to perceive their environment.

q4

What are the requirements for artificial intelligence applications?

Data and storage
We are generating more data today than ever before. For AI applications to produce good results and further improve, they require large amounts of data. But managing large volumes of data is a challenge in itself, especially since data at large organizations often comes from a variety of sources, in different formats and needs to be aggregated. An organization’s data architecture should ideally be adapted to AI.

Computing power
Building high-functioning systems require the right hardware and infrastructure. Improvements in computing power have been key to the progress made in artificial intelligence applications. Faster computers can process more data and perform higher caliber functions as a result.

Algorithms
An algorithm is a mathematical instruction, a step-by-step procedure for calculations. Almost all AI applications are based on algorithms. With the increasing amount of data and computing power, AI experts can create more sophisticated algorithms, which have become so sophisticated that they facilitate machine learning and allow computers to learn on their own.

Q5