New Physics-Based Self-Learning Machines: A More Energy-Efficient Alternative to Artificial Neural Networks
The future of artificial intelligence may be revolutionized by new physics-based self-learning machines developed by scientists at the Max Planck Institute for the Science of Light in Germany. These innovative machines could potentially replace the current artificial neural networks and significantly reduce energy consumption.
Artificial intelligence, while impressive in performance, requires a substantial amount of energy for training. For example, the energy required to train GPT-3, the AI behind ChatGPT, is estimated to be equivalent to the annual energy consumption of 200 German households. This high energy demand has prompted researchers to explore new ways to make AI more energy-efficient.
One promising approach is neuromorphic computing, which differs from artificial neural networks by processing data in a more brain-like manner. The concept of neuromorphic computing aims to mimic the parallel processing capabilities of the human brain, leading to more energy-efficient computations.
The self-learning physical machines developed by Víctor López-Pastor and Florian Marquardt take this concept a step further. These machines optimize their parameters through physical processes, eliminating the need for external feedback during training. This not only saves energy but also reduces computing time, making AI training more efficient.
The researchers are currently working on implementing their self-learning physical machines in optical neuromorphic computers. By harnessing the power of light waves and utilizing reversible, non-linear processes, they aim to create AI systems with significantly more synapses and larger datasets than current neural networks.
In just three years, we may see the first self-learning physical machine in action, paving the way for a new era of energy-efficient artificial intelligence. With the potential to revolutionize AI development, these physics-based machines offer a promising solution to the energy challenges posed by current neural networks.
