The potential of Artificial Intelligence: a very brief introduction.
By André Sekulla, Victoria Wenzelmann, Max Krüger — Universität Siegen
This article first appeared in August 2018 in an internal publication of the Gesellschaft für Internationale Zusammenarbeit (GIZ), the german development cooperation. It is the result of a cooperation between Universität Siegen and the GIZ.
What is AI?
Researchers and technologists have been dreaming of intelligent machines and trying to develop them for several decades. With great increases in computing power and the ability to store and analyse large amounts of data in recent years, Artificial Intelligence (AI) can today achieve exciting results and and its actual applications are now explored on a large scale. At its core AI is the ability of a computer system to perceive specific aspects of its environment and act upon them to reach a predefined goal.
AI technology consists of various different algorithmic methods combined: Data pre-processing methods are used to analyze data and prepare them for the implementation of artificial intelligence algorithms. Following this so-called data mining, statistics-based AI algorithms are applied, which for example cluster items or analyse their significance for a specific context. For the computer system to be regarded as an AI, machine-learning methods such as deep learning are combined with these statistical methods. Machine-learning enables computers to learn: they gradually improve performance on a specific analytical task through repetition, without being explicitly programmed. It is based on multi-layered neural networks, in which each of the individual neurons only contains a simple function — very similar to the human brain and its individual neurons. Their great potential lies in large numbers of neurons forming a network and allowing for an exponential number of combinations. Deep learning AI is able to train and improve itself, thereby growing the number of neurons and possible combinations. This can lead to behaviour that the developers of a system did not anticipate in its design. For example, a deep learning algorithm can find criteria that make an item fit a specific category that its developers did not know about, and perhaps cannot even perceive.
How does AI work?
In order for the procedures of artificial intelligence to be used effectively, they must first be enriched with training data. The training process is varied and important, and generally tries to minimize result errors — for neural networks these errors in need of minimization are calculated by mathematical functions. There are different types of training processes, the best known ones being Supervised, Unsupervised and Reinforcement Learning.
In Supervised Learning , the input values and the corresponding output values are known within the training data. The goal is for the AI to approximate a function — find a relationship between input and output — so well that when it gets new input data, it can relatively reliably predict the output values for that data. An example for Supervised Learning would be training an image recognition AI to distinguish between images of cats and dog: It is known which input data “images” represent the output “cats” and which represent “dogs” — given enough data to learn from, the system will be able to distinguish between new images of cats and dogs. However, what might work well for cats and dogs has proven to be racially biased when applied to humans.
In Unsupervised Learning , on the other hand, only the input values are known, whereas the output values are unknown. The goal here is to find and model underlying structures or distributions in the data which humans would not be able to see — more often than not due to the sheer amount of data, aka “Big Data”. Unsupervised Learning is often used for clustering and data compression: how many images of cats are for example needed to make sure the image is actually one of a cat? Probably not all of the available cat content the Internet offers. Through Unsupervised Learning, an AI can help developers analyse the available images for similarities and assign them to different clusters. In the next step of the learning pipeline, only small samples from each cluster — compressed data — can be used for Supervised Learning.
Reinforcement Learning is very different from the other two methods, as its goal is not to find the correct short-term answer to a given linear problem, but the best solution for tackling a more complex problem in the long term — through trial and error. Reinforcement Learning based methods of AI generate an output based on given input, whereupon they receive feedback and incorporate this into further output generation. To find out what the best solution — or performance option — is, algorithms must find a balance between the exploration of unknown data and the exploitation of known ones. Well known examples for this method are the AlphaGo computer program which beat professional human players in the complex strategy game Go, OpenAI’s robotic hand that taught itself to grip things like humans do. Other applications are stock price prediction and electricity load forecasting. Reinforcement Learning is the method most subjective to its designers’ biases, as it is these humans that teach the algorithm which actions are rewarded and which are sanctioned. It also requires a lot more training data than other methods and is still in most cases outperformed by a combination of human wisdom and simpler algorithmic methods — which makes it the most exciting method with the most resources put into researching it right now.
In practice, it often makes sense to combine different types and algorithms of artificial intelligence. Each algorithm has its individual strengths and often its special use case. In addition, it is very important to know the environment in which artificial intelligence is to be used. The formats of the input and output data are defined before programming the artificial intelligence. Contextual knowledge helps to determine which input data might make sense, which can be omitted and which additional data can serve as enrichment.
Current Ability and Potential of AI
In its current state, Artificial Intelligence can be used to automate specific simple processes, which do not require independent creativity or more complex thought processes.
Programmable AI developed by humans is still a long way from natural intelligence as imagined in Science Fiction, and it is currently not possible to create intelligence that exhibits its own spontaneous emotions. Current algorithms of AI and Machine Learning work on the basis of experience and can only interact based on that. However, AI in various forms of sophistication is increasingly finding its way into many other of application. For example, to detect skin cancer , predict customer behaviour, or as the foundation for robotics — most robots are trained using Reinforcement Learning methods. With concepts like Human-centered AI designers and developers aim to ensure that human needs and abilities are at the center at of AI applications. Ethical issues are increasingly coming to the fore and demand a great deal of attention: Should AI be used in military applications? (It already is) How should the AI in self-driving cars decide in the case of an accident? Should such decisions be made by AI at all? The dependency on the programmers and the training data also makes the automated decisions of AI systems highly susceptible to bias and prejudice such as racism or sexism: the developers’ bias is baked into the decisions of the system. And if collective prejudice such as racism is present in the data a system is trained with, it will recreate that bias in its decisions.
What makes AI relevant for global development?
The vast possibilities of applying AI as well as the associated risks make it highly relevant for development work as well. It is already being used to help small-scale farmers identify plant diseases, assists doctors to diagnose diseases in newborn babies, or helps to save energy. Large technology companies — instrumental in the development of AI — are paying increasing attention to Africa, and invest substantial amounts of money and efforts in the continent: In a program called Lucy IBM research is figuring out how to apply its powerful Watson AI in mostly eastern Africa, and Google is opening an AI research center in Ghana. But not just the beneficial possibilities, also the more systemic risks associated with AI demand the attention of development actors: if the biases of western developers and the associated training data are underlying AI systems these will misalign with local contexts, as the models do not fit the requirements and realities of the application region. Worse, they might increase existing large inequalities by recreating racism, sexism or ableism. In western contexts the fear that AI-based automation will reduce jobs is already large and strongly influences the public debate. In regions where jobs are already scarce, the impact of job loss through automation might be devastating, especially since the jobs that western companies have outsourced to countries of the majority world, such as customer service, textile production or menial software development are often easier to automate, so it seems. Global development organisations need to be mindful of these challenges, counter the risks and apply AI carefully.
Created in cooperation with Deutsche Gesellschaft für Internationale Zusammenarbeit (CC BY-NC-SA 4.0)
