AI at the edge

Welcome to the documentation for the project AI at the edge

Introduction

AI has been on the rise for several years. Not only are today's problems becoming increasingly complex, but the solutions required to solve them are also presenting mankind with ever greater challenges. Computer science plays a central role in the 21st century. Programmers are required to solve problems that have remained unsolved for centuries. This is where the development of AI's comes into play. Since the term AI is very comprehensive and it can be difficult to find the entrance, we have decided to create with the project "AI at the edge" an introduction to this very fascinating topic. The goal of this book is to facilitate the entry into the topic of AI. With explanations, examples and exercises we want to create an opportunity for interested people to deal with the topic. All you need is a Raspberry Pi 3B or better, a mouse, keyboard and a screen.

Intelligence and Learning in Humans and Machines

Before we get to the algorithms of artificial intelligence, we would like to clarify what intelligence means in general. In this chapter we will explain a distinction between strong and weak AI. Furthermore, it is important to know what learning means in general and how machines learn. After we have clarified this, we will set up the Raspberry Pi together and can look at some of the most common algorithms. With the help of exercises, you can also implement what you have learned yourself.

Intelligence

"Intelligence is the best researched characteristic in psychology." (Rost, 2013)

Despite the fact that it is a subject on which so much research is being done, we are still far from understanding it. The term is also controversially debated in education, social science, and brain research. This is the reason why there is no unified definition and the term is considered diluted. In general, intelligence comes from the latin word intellegere (=to see / to understand) and is equated in everyday use with "mental ability". The term refers primarily to the ability of living beings to use the totality of cognitive abilities to solve a problem. The word living being is deliberately chosen because it is not only applicable to humans, but also observed in the animal kingdom (Rost, 2013). According to many intelligence researchers, there is no way to measure intelligence accurately. The IQ test, for example, is criticized as being classist. That means it disadvantages socially low classes and minorities. Others even speak of methodological errors in these tests. But the exact criticism of the intelligence concept would go beyond the scope here. There are numerous theories and approaches that try to describe the cause and effect of intelligence. All of them have their supporters, but also opponents.
One approach is the theory of multiple intelligences according to Howard Gardner from 1983. This theory does not stand up to empirical testing and is therefore widely rejected. Within the academic-psychological intelligence research multiple intelligences are no longer seriously discussed. Nevertheless, it provides interesting approaches which play an important role in artificial intelligence. We will explain why this is so in a moment. According to Wikipedia it understands intelligence as a number of abilities which are necessary to solve problems. This also includes the recognition of these problems. For him, there are 8 intelligences:

  • The linguistic-verbal intelligence includes sensitivity to spoken and written language, the ability to learn languages and to use languages for specific purposes.

  • Logical-mathematical intelligence is the ability to analyze problems logically, perform mathematical operations, and investigate scientific questions.

  • Musical-rhythmic and harmonic intelligence represents the ability to make music, to compose and to have a sense of musical principles such as sound and rhythm.

  • Visual-spatial intelligence includes the theoretical and practical sense of grasping structures and spaces themselves.

  • The bodily-kinesthetic intelligence means to use and control the body and individual body parts precisely. Surgeons and sportsmen possess high physical-kinesthetic intelligence.

  • The naturalistic intelligence includes the ability to observe natural phenomena, to distinguish between them, as well as to develop a sensitivity for them. It also includes the effects of actions on the environment.

  • Interpersonal intelligence means to sense and understand motives, feelings and intentions of other people. It is the essential prerequisite for successful interaction with others.

  • Intrapersonal intelligence is the ability to understand and influence one's own feelings, moods, weaknesses, drives and motives.

("Wikipedia: Theory of multiple intelligences", 2022)

Weak AI

A weak AI has no explicit capabilities to learn on its own or even to perform creative activities. Its learning ability is limited to recognizing patterns and searching large data sets. For this, the tasks must be clearly defined and it must follow a fixed methodology. A weak AI is not able to search or recognize a task independently. It is mainly used for text, image and speech recognition. In addition, translating texts is a classic task. Digital assistance systems such as Alexa, Siri and Google Assistant as well as the Deepl translator are weak AIs (Weak AI, n.d.) Now we see why the concept of multiple intelligences is so interesting. These AIs only operate inside a small part of the previously mentioned intelligences, and even then they are still far behind human capabilities. These voice assistants master a part of the linguistic and the logical-mathematical intelligence. However, they are still very limited in these areas. There are even programs which can independently create pieces in the same style of the composer on the basis of composed classical music. Only for interpersonal and intrapersonal intelligence there are no solutions yet. Some experts even assume that they will never exist.

Strong AI

"Strong AI is the form of artificial intelligence that has the same intellectual skills as, or even surpasses, humans." (Russel & Norving, 2012)

A strong AI can independently identify and define tasks and independently acquire knowledge in order to solve them. The devised solutions can be creative and novel. This AI must use all the previously mentioned intelligences to achieve a goal. It must also be able to make decisions in the face of uncertainty. It is unclear whether this intelligence also needs intrapersonal intelligence, i.e. consciousness and sentience, in order to act logically.

Intelligence measurement for machines

This is at least as difficult as for humans and involves the same complications and criticisms. Consequently, there is not yet a suitable test to accurately measure it. However, there are a number of good approaches. One of them is the Turing Test. Here, a questioner has a conversation with a machine and a human without sight or hearing contact. The test is passed if the questioner does not find out during the conversation which of the two partners is the machine. Now the machine can be assumed to have human-like thinking ability ("Wikipedia: Turing test", 2022). This principle can be extended to painted pictures, composed music and other areas. These Turing-like tests have already been conducted and partially passed. In 2017, researchers at Rutgers University exhibited pictures of an AI at an art fair and the test subjects were asked to guess which ones were painted by humans and which ones were painted by the machines. Overall, the AI paintings were considered more human than the human paintings.
Continuations of the Turing Test are the Lovelace Test and the Metzinger Test. The Lovelace Test asks for evidence of a creative activity. For example, writing an essay according to certain content specifications. The Metzinger Test states that an AI must enter the discussion about artificial consciousness with its own arguments and convincingly argue for its own theory of consciousness. There are still numerous interesting approaches with which one could fill whole books and which invite to philosophize. Therefore we will not go further into it and rather clear up about the misbelief "Games as a Test for Intelligence of Machines".
These tests measure intelligence on the basis of false criteria. An AI may beat the world champion in GO, but the same AI can neither translate texts nor write an essay. So it has only an insular talent. This particular GO AI has been trained with 30 million moves by masters and has also played against itself thousands of times to develop new strategies. So this AI needed an enormous number of games to become so good. Whether this can be called intelligent is questionable. So we see that a variety of tasks is required to evaluate the intelligence of a machine and not just test its insular talent.

How do humans learn?

From a psychological point of view, learning is the process of changing behavior, thinking or feeling based on experience or new insights. This ability is a basic prerequisite for living beings to adapt to the environment and to act meaningfully in it. From a biological point of view, humans learn by linking or by strengthening the connection of synapses. If we do not use the connection of these synapses anymore, the connections become weaker and we forget again. The optimal process of creating and strengthening these connections, i.e. to learn, can be different depending on the individuals. Some learn better with text, others need to hear something, and still others learn better through pictures and videos.

How do machines learn?

"A computer program is said to learn from experience E with respect to some class of tasks T and performance measure P, if its performance at tasks in T, as measured by P, improves with experience E." (Tom Mitchelle, 1997)

The programs have a task, know how to evaluate their performance on that task, and are then said to improve performance from the experience gained. If they have improved performance from the experience gained, then they have learned. The experience also includes data with which the program has already been fed beforehand. This data comes in different forms. In supervised learning, the program is given the correct value for each example. These are then so-called labeled examples. Here it must be said that the labeling of these examples is very time and cost intensive. This includes decision and regression trees as well as classification rules. Then there is semi-supervised learning, where only a part of the training examples is labeled. In reinforcement learning the program receives a reward by interacting with its environment at certain points in time. This reward can be positive or negative. The goal is to maximize the rewards. Last but not least there is Unsupervised Learning. Here there is no information except the training data itself, so there are no target values or rewards. Algorithms in this category are for example Association Rules and Clustering Techniques.

References

[1] Rost (2013). Handbuch Intelligenz S. 11
[2] Wikipedia: Theory of multiple intelligences (last accessed on 06.01.2022)
[3] Schwache KI. (o. D.). Hochschule für angewandte Wissenschaften Würzburg-Schweinfurt (last accessed on 06.01.2022)
[4] Künstliche Intelligenz ein moderner Ansatz, 3 aktualisierte Auflage, Stuart Russell, Peter Norvig, 2012 (Russel & Norving, 2012)
[5] Wikipedia: Turing test (last accessed on 06.01.2022)

Written by Kai-Philipp Nosper and Kevin de Riese-Meyer