Artificial intelligence is convenient: ask a quick question and get a detailed answer almost instantly, without having to sift through numerous search results. Yet what is convenient is not necessarily energy-efficient: while a traditional Google search for a restaurant may consume roughly the same amount of energy needed to power a light bulb for a few seconds, an AI model’s response to the same query can require up to 20 times as much energy. The reason lies in the underlying computational process: AI models are enormous, and every query triggers a large number of calculations that must then be interpreted and converted into coherent text. This process consumes a huge amount of computing power and, consequently, large amounts of electricity.
Another factor contributing to energy consumption, whether in search engines or AI systems, is the location of the data. The data and AI models are typically stored in large data centres, often located far from the user’s computer. Data from the query must therefore first be transmitted to these facilities, processed and then sent back. And all in a matter of seconds.
Prof. Yulia Sandamirskaya, who heads up the research area at the Institute of Computational Life Sciences at the ZHAW, offers a fitting comparison: “You can think of this process as a parcel being routed through a postal distribution centre before reaching the recipient, even though the recipient lives in the neighbouring village.”
Yulia Sandamirskaya
Research Area Head, Institute of Computational Life Sciences, ZHAW School of Life Sciences and Facility Management
“We need to use new technologies to become more energy-efficient and give people back control over their data.”
Local data for robotics
Together with her team and various partners, Sandamirskaya is researching a new type of technology: neuromorphic computing chips. The term “neuro” within the name already suggests that the chips might have something to do with the human brain. And that is exactly the idea: Yulia’s research group is attempting to emulate biological neural networks using algorithms and to apply them in a new generation of computing chips. These could be applied in the healthcare sector, signal processing or in robotics, for example in agriculture or in laboratory automation.
One of their key innovations is that the chips would store and process data locally instead of transmitting it to large data centres somewhere far away. In metaphorical terms, the parcel I just mentioned would simply need to be taken into the next room. And that would save enormous amounts of energy. “The data should remain where it is needed,” says Yulia. This is especially important in robotics, as speed is of the essence there. While artificial intelligence generates an output from an input, or prompt, a robot’s movements are based on numerous sensor signals that are fed into its motion control system. The aim is to process these sensor signals in real time using neuromorphic chips.
Neuromorphic chips as the antithesis of AI
A chip that functions in a manner similar to the human brain raises an obvious question: how can such a system be recreated? To answer this, the research group works closely with various disciplines within neurobiology and cognitive science to gain a better understanding of how biological brains function. “We still know surprisingly little about them,” says Sandamirskaya. She even advocates for more discussion in schools about the characteristics of our brains in order to raise awareness among children and young people about mental health challenges, neurodivergence, and their own learning processes and emotional responses.
The Professor of Cognitive Technology takes a critical view of current approaches to artificial intelligence: “As AI is built today, facts are often jumbled up, causing people to become mentally fatigued, and the development and maintenance of their own cognitive abilities are neglected. In addition, the reckless sharing of sensitive data poses major risks if it falls into the wrong hands.”
Safety considered from the start
Many of the shortcomings of today’s AI systems become particularly apparent when they are used in robotics. Safety is perhaps the most important example. Sandamirskaya explains: “Robots are machines. And with machines, safety is always considered and certified.” She compares this to cars, which have to undergo countless safety tests before entering the market. When it comes to software, including artificial intelligence and social media, these certification processes do not currently exist. Prof. Yulia Sandamirskaya considers this to be a deficiency. “It is high time for software to be regulated as well and for developers to be required to comply with safety standards.”
Yulia’s vision with neuromorphic chips is clear: to make the market more open and broader. Data should be stored and accessed locally, wherever it is needed, rather than being concentrated in the hands of a few dominant players with vast server infrastructures that place a burden on the environment. “We need to use new technologies to become more energy-efficient and give people back control over their data.”
