AI-based predictions — A blessing or a curse?
A scientific theory works when it makes reliable predictions about future events or conditions. It involves building a model of the world, deriving predictions from that model and testing those predictions. Understanding why something is happening is different story. And while AI may never be able to understand the “why”, what it can do very well is deal with data.
AI is often equated with GPT and other generative models, but the recent successes of machine learning and AI are far more significant than the hype surrounding GPT. After all, the secret of the progress made by language models, which even experts deem amazing, is nothing other than their predictive power. These indescribably large neural networks have not actually cracked the “problem” of language. They are simply much better than their predecessors at guessing the most likely continuation of a sentence. What amazes us so much about the result is “only” the prediction of the next word.
Language models also learn from images
“Finding the right next word – that’s a pretty difficult task,” counters Robert West, Assistant Professor of Computer Science at EPFL, where he heads the Data Science Lab (dlab) and researches these kinds of language models. “It takes a little understanding of the world to do this well. ” This leads us to the central issue of the current AI debate: are language models more than just models for language? Recently, these models have been learning not only language or images, but everything at once. “The image input then improves the language system. One model reinforces the other,” says West. These “multimodal” models could then actually learn things simply by looking at them. As an example, West cites the simple fact that “if you let go of something, it falls to the ground”.
What would happen if robots were equipped with such systems? “Human intelligence is anchored in the physical world,” says West. According to him, children are really “little researchers”, and people learn all the time through trial and error; machines could do that too. Artificial intelligence can already do a great deal of work for human scientists.
“Language of nature“ as a role model
Modelling, testing and validation: Siddhartha Mishra, a professor at ETH Zurich, knows the three-pronged scientific approach well. The mathematician and machine learning expert has been working on teaching machines this type of analysis for a long time. Mishra specialises in partial differential equations, a mathematical tool that can be used to describe many natural processes. He calls them the “language of nature”. However, working with this language is much more complex than what ChatGPT and similar tools can handle. Human language is ultimately a highly compressed version of the world. “Modelling natural systems is a far greater challenge due to their inherent complexity.”
But, according to Mishra, we already have a much greater knowledge of nature, which we simply have to incorporate into the models, adding that previously, solving partial differential equations would have necessitated the use of a supercomputer due to the computational demands. AI now makes such problems considerably easier to manage. Mishra collaborated with Empa researchers to develop fast algorithms for simulating laser manufacturing processes for 3D printing. Thanks to AI, the simulation time has been reduced from four hours to microseconds, enabling real-time simulation and printing. The use of partial differential equations is also common in tsunami early warning systems. Mishra’s approach could be lifesaving if it can quickly calculate the simulation of expected events following an earthquake.
Energy policy scenarios
Philipp Heer, who is responsible for AI in the construction and energy sector at Empa’s Digital Hub (dhub), explains that AI is used to solve classic optimisation problems, such as efficient house heating. Additionally, AI is used to calculate energy policy scenarios for Switzerland. However, issues of trust arise when machines dictate future policy decisions. Heer argues that AI generally performs tasks in a similar way to human experts, but at a much faster rate. “The outcome would be comparable to that of 10 researchers working on the problem for 10 years.” According to Heer, the current systems allow for the modelling of a significantly greater number of scenarios.
Nowcasting urban flooding
João P. Leitão, group leader in the Urban Water Management Department at Eawag, does not speak of ”forecasting” in this context, but rather of ”nowcasting”. Forecasting urban flooding events after heavy rainfall is only possible through the use of AI models. Previously, computers would have had to be run for hours in order to obtain similarly high-resolution simulations - any warnings would then come too late.
This raises thoughts of science fiction. Will it be possible to simulate everything in the near future? Will AI be able to calculate any possible data sequence accurately? Mishra says that physical modelling has not yet paralleled ChatGPT’s trajectory, meaning that there has been no magic moment yet. However, with the current pace of progress, it is difficult to predict where things are headed. Regardless, Mishra remains hopeful and optimistic: “My main research goal is to develop an AI that can model nature as effortlessly as GPT models language.”
West also believes that current models can do much more than generate language outputs: “They are used as next-word predictors. However, this prediction is not necessarily limited to words.” With the appropriate data, it can also work with credit card transactions, online behaviour and even gene sequences.
Prediction of plankton growth
Models and success stories are ubiquitous. With this in mind, Marco Baity-Jesi takes a more grounded view of current developments. As a physicist, he researches both applied and theoretical AI approaches at Eawag, such as simulating plankton growth in lakes. The benefits of AI are evident: it can handle vast amounts of data, model complex systems even with incomplete information and make accurate predictions. However, this strength also calls for caution, as the future is inherently uncertain, and predictions are only accurate to a limited extent. “We have a tendency to fall in love with our models all too easily,” he says. “And validating the results is often not so straightforward.” A lack of critical thinking in the AI community could be a concern, he believes, and it is essential to evaluate the reliability of predictions.
Recognition of systematic errors
Konrad Bogner, a WSL hydrologist, is familiar with the phenomenon. An old hand in the field of AI predictions, he worked with neural networks for his doctoral thesis 20 years ago to model run-off volumes after storms. Although hardly anyone had heard of the term back then, today everyone is trying AI. While the buzzword makes it easier to sell scientific work, it is important to consider whether there is an actual benefit at all. While simpler means may suffice for some fields of application, a differentiated approach can offer significant potential, as Bogner suggests. For instance, AI modelling can help identify systematic errors in predictions, a phenomenon known as “bias” in language models. These models are not neutral and can adopt our biases, and the same is true in the fields of meteorology in meteorology and hydrology. Highlighting the weaknesses in conventional models can help us to spread awareness of these weaknesses.
Mishra argues that models that simulate natural phenomena have an advantage due to their “groundedness”, as opposed to models like ChatGPT that attempt to master the ambiguous concept of language. “In natural phenomena, ‘truth’ is well-defined,” as evidenced by the accuracy of weather forecasts. We have extensive experience in testing model accuracy related to natural phenomena, but what about in the context of medical diagnoses? When asked if he would trust a machine-assisted system in this field, Mishra did not hesitate to answer affirmatively: “I would trust it without giving it a second thought.” His wife is a doctor, and he feels certain that she and her colleagues also use machine learning. “Many people don’t realise that doctors also work with models, constantly approximating the truth.” Therefore, the combination of human and algorithmic approaches makes sense to him.
Early detection of cancer
Researchers at the PSI are currently exploring what this could look like in practice. G.V. Shivashankar, Head of the Laboratory of Nanoscale Biology at the PSI and Professor of Mechano-Genomics at ETH Zurich, uses image recognition AI to elicit medically relevant information from specific blood cells.
What was actually developed to identify objects in photos helps to recognise the ageing processes in cells or to enable the early detection of cancer. The researchers are taking advantage of the fact that DNA is “packaged” differently depending on the state of a cell, which can be visualised using appropriately sensitive microscopy techniques.
With the help of semi-supervised learning (hybrid technology that uses labelled and unlabelled data), the AI can be trained to recognise level while also acknowledging blood cells in a body in which a tumour is growing with a high degree of accuracy. “It has always been our underlying assumption that every cell is important in understanding disease,” explains Shivashankar. However, tracking clinically relevant changes at the level of individual cells is very challenging and can only be achieved with AI. He considers the current processes to be “so powerful” and their application in medicine so beneficial that completely new AI methods will emerge from this synergy. He is convinced that diagnostics at an early stage of the disease will become far more important than it is today.
Man-machine
What is clear is that the relationship between humans and machines will undergo a significant shift in the coming years. Heer notes that this shift has already been felt in the building sector, where the automation of heating systems has highlighted our “desire to control our environment”. People are suspicious of automatic, self-regulating systems.
Leitão also believes that AI-supported processes will require human control. The way AI systems are currently built, they lack the crucial human capacity for decision-making: "The concrete decisions always have to be made by humans."
West believes that political regulation is necessary to renegotiate this relationship. He commends current efforts at the EU level while also acknowledging the potential of AI methods, particularly in the medical field. However, regulating AI is not the same as regulating industrial robots or a steam engine due to the models’ ability to improve themselves: “They can explore the real world to understand how it works.” It is currently challenging to predict whether this will result in a significant improvement in the models. “If such a development occurs, humans may no longer be able to make predictions” about the future of AI.
For West, one thing is certain: “Something big is coming.” ChatGPT has made the general public aware of this. “Language is so obvious a choice because it’s so human, and because so much can be expressed in language. But it’s not going to stop there.”
