A research team from the Technical University of Munich has developed a new method for training artificial neural networks without the need for energy- and resource-intensive computing.
According to results presented at the Neuro-IPS 2023 conference and in a newly published study , this new method does not rely on traditional optimization methods, but rather adopts a novel approach that makes it approximately 100 times faster and, therefore, more energy efficient.
"This approach can be considered a radical change, but it's still in its infancy," Mohamed Khalif, a professor of computer science at the Arab Academy for Science, Technology and Maritime Transport in Egypt, told Al Jazeera Net exclusively. "It's like someone who understands and knows where to look to find the answer, relying on intelligent guesswork. It could represent a shift from relying on the quantity of data to focusing on the quality and intelligence of the learning process."
This new method does not rely on traditional optimization methods, but rather follows a new approach (Shutterstock)
expensive computer brain
AI applications like ChatGPT, Gemini, and other large language models have become an integral part of our daily lives, and data centers—which consume massive amounts of energy—provide the computing, storage, and transmission capabilities needed to run these models.
In Germany alone, this consumption reached approximately 16 billion kilowatt-hours in 2020, equivalent to nearly 1% of the country's total energy consumption, and this figure is expected to reach 22 billion kilowatt-hours this year.
Neural networks are the foundation of modern AI tasks such as machine translation, image recognition, and large language models. They operate through a massive collection of interconnected artificial "nodes" or "neurons," where signals are passed through and modified based on specific "weights."
