Since 2008, the second wave of AI applications has enabled surprising new functions, raising expectations so high that today every currywurst stand claims to have an AI strategy. The models driving this technological revolution rely on huge artificial neural networks (deep learning) that fit millions (and later billions) of parameters using carefully crafted training datasets. But until recently, all these AI models had one thing in common: They were limited to processing specific inputs (like camera images from vehicles) with narrowly defined results (like pedestrian detection) upfront. The basis are human-defined goals using manually annotated training data (supervised learning). The behavior of such models is easy to understand and measure, and their performance is easily quantifiable – such as in a confusion matrix that indicates how many real pedestrians the artificial intelligence (AI) has missed or thinks it has recognized where there are none.
Jonas Andrulis, Founder and CEO Aleph Alpha
Aleph Alpha GmbH, founded in Heidelberg in 2019 by Jonas Andrulis and Samuel Weinbach, is the only European AI company to conduct research, development and design in the direction of generalized artificial intelligence (Artificial General Intelligence, or AGI for short).
The Heidelberg team is contributing the latest generation of multimodal models to the Gaia-X project OpenGPT-X. In 2021, they set a German deep-tech financing record with 28.3 million euros in venture capital.
Technology sovereignty creates the ability to act
According to the company, it is self-confidently striving to bundle technical competence and added value here so that Europe remains able to act in global competition. "If the entire value creation goes to the shareholders of Microsoft and OpenAI, then we lack that here as a society," CEO Andrulis underpins the concern. The machine learner and serial entrepreneur was previously active in leading AI research at Apple. In 2021 he won the German AI Prize for a technology comparable to Deep Mind and OpenAI.
AI in an expert discussion: Where is Europe headed?
In the Heise interview "Free kick or own goal – more scope for machine learning", Jonas Andrulis and Reinhard Karger (DFKI) discuss artificial intelligence and technological transformation in a complementary manner.
When building this type of AI, it's possible to accidentally train a model that doesn't perform well for a particular group of people or in certain use cases. Thus, once such a potential problem is identified, it is relatively easy to also raise ethical concerns about the capabilities of the model: pedestrians (in this example) with a rare appearance (meaning here: unusual compared to the majority in the dataset) might not have been recognized with sufficient quality. If this is the case, it can be clearly evaluated and presented.
What is missing: In the fast-paced world of technology, there is often the time to re-sort all the news and background information. At the weekend we want to take it, follow the side paths away from the current, try different perspectives and make nuances audible.
We can attempt to create a (partial) dataset of pedestrians with unusual appearances (based on appropriate criteria such as ethnicity or size) and measure how well our model performs on this subgroup. If the result is below our acceptance threshold, we can add more training data that contains exactly those observations or increase the weight of the ones we already have. This approach won't make the models perfect, and trying to remove every unwanted behavior will lead us into an endless loop of changes – but at least the tools are well established.
Prominent examples where these types of systems have sparked debate include the misclassification of humans as monkeys, or the poor performance of facial recognition for black users. In these examples, very clear and well-defined goals for the AI system (recognition of objects in images, recognition of human faces) were undoubtedly not achieved for a subset of the data. While no AI system will always get everything right, it makes sense to aim for a system that avoids these types of mistakes. The fact that these cases were promptly corrected by the companies concerned demonstrates the possibility of subsequent improvement. With her, one should be aware that trying to fix these issues is an iterative process, and systems must allow for repeated adjustments along the way. There are now proven tools and methods to reduce such unwanted effects while maintaining the basic functionality in a similar quality.
A new generation of world models
A new generation of AI models has existed in research and industry for several months, for which similar evaluations can no longer be carried out in a trivial way. These models do not attempt to learn a specific monitored (predefined) association (camera image on pedestrian) but to understand general "meaning" and patterns in the data. Since human annotation is no longer necessary, this procedure is called self-supervised learning. With a sufficiently large model and data set, it can and will find complex structures and dependencies. Based on these models and their learned world structure, a variety of different use cases can be implemented that leverage the knowledge gained during training.
For this reason, a group of researchers at Stanford University have given them the name "Foundation Models" – because with their help we can develop countless new possible applications that build on the world knowledge of the models. Many of the skills that emerge from this build are surprising and were not foreseen or planned during training. The results are often novel and can be impressively complex: the DALL-E generative image model developed by OpenAI, for example, can create images of the desired content based on a short text description, with almost unlimited flexibility (see Fig. 1).
DALL-E from OpenAI generates images based on a description of the desired image content (caption) (Fig. 1).
The functionalities that can be built on these basic models are almost unlimited – they use complex and sometimes unknown internal patterns and relationships in the data (and the world). As a result, there are no easy measures to address ethical concerns similar to supervised learning models. If DALL-E creates an "armchair in the shape of an avocado", how can we judge whether this depiction is fair? Are all types of chairs or environments represented as we deem necessary? Are people included? What are our requirements for depicting people? While these questions may seem harmless for the DALL-E example, they become extremely relevant in the context of GPT-3 and similar models.
For example, GPT-3-like models are able to write a summary of a long piece of text without having to see examples or additional training data. But will this summary adequately reflect all aspects of the document? Wouldn't a psychiatrist write a different synopsis than an engineer, or would an old German concentrate on different aspects of the text than a young Japanese writer? Is this subjectivity, which we take for granted in human authors, also okay for AI models? How can this be measured and compared fairly?
We need to examine two questions here:
What do the basic models actually do? How can their results be understood? What are reasonable requirements for these models from an ethical perspective?