With guest blogger Patrick Klingler*
Artificial Intelligence is one of the hottest topics of 2018 and thousands of articles have appeared on the topic. Academics, consultants and others have been forecasting the impact of AI on society, the business world and even private life. There are so many perspectives on AI, which is absolutely okay given that the topic is very broad and involves several scientific and social-scientific disciplines. However, a lot of authors fail to clearly define AI when writing an article, and this often leads to difficulties and misunderstandings for the reader.
In this blogost we want to help you to tackle this issue by providing guidelines to (i) understand AI articles and (ii) be able to develop a critical view.
An approach to tackle AI
An increase in processing power and the availability of huge amounts of data has led to the renaissance of machine learning (ML) based AI. So in most writings about AI, the author refers either explicitly or implicitly to machine learning or deep learning (DL). However, we want to stress that ML is the mostly often references, but not the only dimension of AI. But what is ML and how should we to structure the AI-related terminology which is often used?
The classical approach to build computer systems is to define a deterministic program sequence in the source code. To do so, you can use operations like if-then-else statements or for-loops to process data and calculate the desired output. In contrast, a ML-based system is created based on data and training. You select a ML-algorithm, feed it with training data, interpret the output and refine the configuration. That process has to be repeated until the outcome reaches a desired quality-level. So the mantra of a classical programmable system is “program and deploy” whereas the mantra of a ML-system is “train and optimize”.
Also, it definitely helps to distinguish between four layers of AI to structure terminology:
- Big data: as mentioned above, data is the base of ML-based AI. Therefore you need big amounts of high-quality data to train ML-algorithms.
- Machine learning: ML is the layer that brings in the “intelligence”. It makes use of historic data to identify patterns. Those patterns are applied to unknown data to derive probabilistic predictions. DL can be considered as a sub-discipline of ML that makes use of artificial neural networks with a large number of layers for data processing.
- Cognitive computing: cognitive computing can be seen as a symptom of AI. It is about the imitation of human cognitive abilities like speaking, hearing or seeing.
- Rational decision making: at the end of the day, every ML-based AI system has the objective to support the rational decision-making process. That’s a logical fact as the system follows a strict rational statistical approach.
Now you are familiar with the basic tool set to understand the concept of AI on a very high level. As a second step, it’s helpful to think about some simple questions that help to categorize and challenge an AI article, opinion piece or blog:
Does the author actually write about ML-based AI?
Besides ML there are other techniques for creating “intelligent” agents. For example, robots that perform pre-defined actions based on sensory data or decision-making based on pre-defined decision trees could also be defined as an AI discipline. To be able to follow the author’s thoughts you should find out which technique he or she is writing about. Sometimes this question can cause problems, as the author fails to clearly define the personal definition of AI. Please keep in mind that this guideline is focused only on ML-based AI.
How does the author define “intelligent” behavior?
Generally, there are two ways in which AI-enabled “intelligent” agents can behave. (i) In a pragmatic way “intelligent” agents behave rationally as their decisions are based on data and statistical techniques. So they are only provided with specific domain knowledge and are not able to develop a generic model of the real world. (ii) Second, the behavior of “intelligent” agents can be compared to human behavior. Sometimes human behavior is equal to rational behavior, but often humans follow their own individual (sometimes un-rational) rules in decision making. Also, humans are very good at generalizing and transferring concepts.
While interpretation (i) can be challenged in a very objective way, interpretation (ii) usually targets a subjective Philosophical debate which gives room for different opinions. In some cases the background of the author can provide an indicator how “intelligent” behavior can be interpreted. A scientific author usually tends to interpretation (i) whereas a social-scientific author tends to interpretation (ii).
Is the funnel from data to decision-making fully transparent?
A well-functioning ML-based AI-system has to go through the process of data provisioning, training of the algorithm and creating probabilistic predictions. You should check if that funnel can be re-engineered in a transparent way, or if AI is simply replaced by a magical black box. Quite simply, there are some situations where ML can’t be applied: (i) the problem can’t be approached by analyzing historic data, (ii) the required data is not available in quality or quantity, (iii) the problem is very dynamic, so data-validity is frequently changing. If you can find contradictions or inconsistencies concerning this point you should be critical about the author’s statements.
Never stop challenging
We hope this short guideline helps to de-mystify AI and to make you re-think some of the stuff that you’ve read. We recommend that you always try to fully understand and challenge the assumptions and statements being presented in an article, blog or opinion piece before you act upon the arguments presented. This will help you to avoid being fooled by unfounded (and often very scary or excessive) stories about the future of AI.
Of course, we don’t want to exclude our blogpost from this advice. So, we’re very happy to read your experiences and opinions about ways to make sense of what is being written about AI. The field of AI is evolving rapidly, so please feel free to share our blogpost and express your thoughts!
*Patrick Klingler is a thought leader in the field of artificial intelligence, and researches IT Innovation Management as part of his role at Daimler AG. LinkedIn: https://www.linkedin.com/in/patrick-klingler-682a6681/