The dream of a scientific revolution
In this post I think about artificial intelligence (AI) and the communication of its achievements. (That's why I've linked many AI blogs here. But there are thousands of them.) Why is artificial intelligence and its deep learning field exciting for a wide audience? Has artificial intelligence also contributed to an increasing enthusiasm for natural sciences? And if so, why? I believe an important answer to this question is: open science.
In elementary school in socialist Czechoslovakia, we children were drummed into Lenin's saying “learn, learn, learn” every day. This compulsion to learn could only have one consequence: we hated learning like cod liver oil. Lenin as well. Lenin still hasn't become my role model, but learning quickly became one of my favorite hobbies. Unfortunately, there were almost no non-fiction books in the city library of my small Moravian home town, except for the Marxist-Leninist ones. That's how I learned science from novels – from science fiction.
During my studies at the Technical University of Munich, I read many non-fiction books on quantum mechanics and biographies of famous quantum physicists. How nice it would be, I often thought, to experience such a scientific revolution: like the quantum mechanical one in the 1920s and 1930s. Dreams often fight against their fulfillment, but my dream was fulfilled:
For a few years now, we have been in the midst of one such scientific revolution: the revolution in artificial intelligence (AI), particularly in its deep learning domain. Only today AI experts earn a lot more money than quantum physicists did in the first half of the 20th century.
Every day I get tons of AI newsletters from all over the world (Figure 1) – it's an intellectual rush like no other. But what is even nicer: In contrast to the dawn of quantum mechanics in the 1920s, everyone can take part in this scientific revolution. Thanks to the Internet, we all have access to limitless information about the area. This popularization of artificial intelligence also makes us more curious about other natural sciences and technologies. I want to show that in this post.
Figure 1: My Outlook with some AI newsletters.
Explanation of terms: artificial intelligence, machine learning, deep learning
But first, I'll explain how artificial intelligence, machine learning, and deep learning are related. What exactly do these terms mean? It has to be: Artificial intelligence is still confused with Terminator. Thankfully, most of us are no longer afraid of this wonderful technology. Since the computer science pioneer John McCarthy coined the term artificial intelligence in 1955, we have become quite accustomed to it and its name. According to McCarthy, “Artificial Intelligence is the science and engineering of making intelligent machines”.
Figure 2: Explanation of terms.
Artificial intelligence is a generic term for programs or machines that can mimic all aspects of human thought and behavior. A distinction is made between weak artificial intelligence and strong or general artificial intelligence:
General artificial intelligence should at least be on a par with human intelligence, be able to react sensibly to its environment and assess the consequences of its actions. Also, it would have to have something resembling common sense and consciousness. (Although we humans cannot agree on what our consciousness is.) Such artificial intelligence does not exist until now. We are in the age of weak artificial intelligence: AI is also a program that we have taught through some kind of training to sort cucumbers according to their quality.
There are two main streams in AI research: symbolic AI and neural AI. Neural AI – artificial neural networks (ANN) – is making most of the headlines today. If ANN have more than one hidden layer of neurons, they are called deep learning. Your area is then called Deep Learning. ANN are also the subdivision of machine learning:
Machine learning includes programs that use many examples to learn how to complete tasks without being explicitly programmed to do so. These can also be classic machine learning programs such as decision trees or support vector machines. But the neural AI (deep learning) plays the most music here – as I said – when we hear or read something about artificial intelligence in the media today, it usually means deep learning.
Artificial intelligence and open science (Open Science)
But how is artificial intelligence supposed to have helped to get a broad audience enthusiastic about science? By introducing AI to the general public more and more openly since 2012.
The natural sciences developed partly from the secret sciences such as alchemy – there were no others in the Middle Ages. However, there is still a hint of occult science in many scientific articles: the results of your study are only reported to a circle of insiders: in a cryptic language. Some of us have laboriously acquired these through long studies and doctorates and through reading thousands of incomprehensible scientific articles. I too am a shining example of this. Below you can see one of my scientific articles 😊 (fortunately from the 1990s):
Figure 2: My scientific gibberish. Or does anyone understand the summary (abstract) of the article? 🙂
Luckily that is changing. In my opinion, also thanks to artificial intelligence (AI), but only in a figurative sense: namely through the enthusiasm for this old-new science and technology of artificial intelligence. In the last ten years, AI has not only been developed at universities, but also by online platforms and tech companies: Google, Facebook, Amazon, NVIDIA, Microsoft, IBM and others have recognized that they can only sell us something if we are enthusiastic about it . And if they act openly: Open Science. (Their AI blogs are linked directly to the platform names.)
Every institute, every university, every platform that wants to be noticed now maintains a generally understandable scientific blog. And not just in data science — in all sciences. MIT Technology Review is exemplary here.
Deep Learning Revolution
English-language blogs about artificial intelligence and its real existing field of deep learning inspire inquisitive people all over the world. Especially neural AI – deep learning – has been attracting researchers since the victory of the deep learning model AlexNet in the ImageNet challenge in image recognition in 2012 and is constantly motivating new AI blogs. (Until 2017, the ImageNet challenge – ImageNet Large Scale Visual Recognition Challenge – was held at the largest image database in the world, ImageNet. After 2017, the challenge was moved to the data science platform Kaggle.)
The SciLogs of the Science Spectrum and Science Spectrum itself have been providing us with understandable science in German for years. But the AI carousel is also turning faster and faster in German-speaking countries. Artificial intelligence is occupying more and more of our knowledge playground:
In 2016, for the first time, a deep learning model – AlphaGo from Google company Deep Mind – beat one of the world's best players of the Chinese game Go – Lee Sedol. This machine victory experienced a wide media interest. Since then, there has been an increasing search for artificial intelligence. If this search is now stagnating somewhat, it is certainly due to the fact that more and more related terms are also playing a role here: not only “artificial intelligence”, but also “machine learning”, “deep learning” and “artificial neural networks ”.
Figure 3: Google Trends for searches for “Artificial Intelligence” between 2004 and 2022. In 2016 (blue dot), Korean gamer Lee Sedol was defeated by the deep learning model AlphaGo in Go.
In the new data science, to which artificial intelligence can be counted, it was quickly recognized that if science is to exist as the only world view that is ONLY supported by facts, it must be understood. Or is there another communication system than that of the natural sciences that we can agree on to some extent around the world? If we put something on reason? Regardless of the form of government? Of course, only if there are no conspiracy believers and alternative-fact preachers at the levers of powerful states.
Other forms are also increasingly popularizing the natural sciences: Open Science is the buzzword: science slams, science cabaret, science cafés. The academies are becoming more open, their formats more entertaining. For years, on behalf of the German Academy of Science and Technology (acatech), I have been allowed to "bleed off" their writings, to make fun of acatech and its writers.
The general director of the Deutsches Museum and holder of the chair for science communication at the TU Munich Wolfgang M. Heckl, Marc-Denis Weitze from acatech and the freelance journalist Wolfgang Chr. Goede have published the pretty anthology “Can science be funny?” brought out with the announcement: “'Can science be funny?' takes a close look at an element of modern science communication that is as innovative as it is promising: comedy!”
Blog and science sites like Medium, Towards Data Science, Quanta Magazine, and more are wonderful collections of science blog posts. Especially in the field of artificial intelligence and data science.
Figure 4: The Medium blog platform
The open science contributions of scientists, science journalists and bloggers inspire more and more young people for the world of artificial intelligence. Hand in hand with this interest in AI, however, is the interest in natural sciences as a whole. The corona pandemic certainly contributed to this: before the pandemic, the profession of AI expert or data scientist was considered one of the most attractive. After two years of the pandemic, the job of a virologist is also desirable. Of course, I am well aware that pseudoscience and anti-science are becoming more and more visible. But in my opinion they were always there, just not as loud as today.
AI crime novels
In the pilot episode "Black and White" I get to the bottom of a rather shocking AI conundrum: The international study "Reading Race: AI Recognises Patient's Racial Identity In Medical Images" found that "Artificial Intelligence" models recognize blacks from white patients can only tell from the x-rays of their internal organs and bones. However, the researchers in the study rightly fear that this could lead to biased decisions in medicine: what happens if an AI model classifies skin color instead of cancerous tumors? No radiologist can tell the color of the patient's skin from the X-ray of the lungs. But why can an AI model do that? What is the reason for that? Of course I try to crack the riddle myself.
The second episode of the AI crime series “LIDAR or not LIDAR” will be published in the “KI Krimis” channel on February 28th, 2022. It is about self-driving cars and whether only cameras are sufficient for the perception of their AI models. In the third episode I ask: "Can there be strong artificial intelligence that would be equal to human?"
Pioneering Deep Learning Enlightenment
Figure 5: Yannic Kilcher's YouTube channel.
In one of his recent posts, Yannic Kilcher reviews an article by Facebook AI researchers (Meta AI): “CM3: A Causal Masked Multimodal Model of the Internet.” They taught their Transformer model CM3 to process actionable information from the Internet.
That's nothing new, you probably think, we also know that from the OpenAIs Transformer model GPT-3. But this CM3 model is trained directly on the HTML code: the language of the websites. Thus, the model not only learns from the text and images simultaneously, but also learns the structure of the processed websites. Therefore, after training, the model can solve complex tasks: a wide range of text and image tasks and many other cross-modal tasks.